.env

@@ -102,10 +102,10 @@ DOCUMENT_PARSE_ARTIFACT_PREFIX=artifacts
 PARSER_FAILURE_MODE=fail
 
 # ===== Reranker 配置 =====
-RERANKER_ENABLED=true
+RERANKER_ENABLED=false
 RERANKER_BASE_URL=http://6.86.80.4:30080/v1
 RERANKER_MODEL=BAAI/bge-reranker-v2-m3
-RERANKER_API_KEY=
+RERANKER_API_KEY=sk-fVr9KmDZNC4pGDBQj0EUWz9bDmFzNxjYC9EzZpe2bVDsxtz8
 RERANKER_TOP_K=5
 
 # ===== 会话持久化 =====
@@ -120,3 +120,10 @@ AUTH_ENABLED=true
 
 # ===== CORS =====
 CORS_ALLOW_ORIGINS=http://localhost:5173
+
+# ===== HyDE ???? =====
+HYDE_ENABLED=true
+HYDE_MAX_TOKENS=200
+HYDE_LLM_PROVIDER=qwen
+HYDE_LLM_MODEL=qwen3.5-flash
+

.env.example

@@ -138,6 +138,31 @@ AUTH_TOKEN_EXPIRE_MINUTES=480
 # 设为 false 可跳过认证（仅限本地开发调试，生产必须 true）
 AUTH_ENABLED=true
 
+
+# ===== HyDE 查询增强 =====
+# HyDE (Hypothetical Document Embeddings): 在检索前让 LLM 生成一段"假设性回答"，
+# 用该段落的 embedding 代替原始查询 embedding 进行向量检索。
+# 无需新模型，复用现有 LLM 和 Embedding 服务。降低此功能可减少每次查询的 LLM 调用次数。
+HYDE_ENABLED=true
+HYDE_MAX_TOKENS=200
+# ?????? LLM;???????????????
+HYDE_LLM_PROVIDER=qwen
+HYDE_LLM_MODEL=qwen3.5-flash
+
+# ===== Agentic RAG 配置 (P0-1) =====
+# 以下参数控制 /api/v1/agent/agentic/stream 多步推理管线
+# 意图分类: simple_qa / compare / multi_hop / ambiguous
+# compare 和 multi_hop 触发查询分解，最多 AGENTIC_MAX_SUB_QUERIES 个子查询
+AGENTIC_MAX_SUB_QUERIES=4
+# 引文锚定 fast-path 阈值: avg_score > 此值 且 chunks >= 3 时跳过 LLM grounding check
+# 降低此值可让更多查询触发 LLM 二次验证（更准确，但延迟+成本增加）
+AGENTIC_GROUNDING_THRESHOLD=0.65
+# 各步骤 LLM 最大 token 数（越小越快，越大越准）
+AGENTIC_INTENT_MAX_TOKENS=200
+AGENTIC_PLAN_MAX_TOKENS=400
+AGENTIC_GROUNDING_MAX_TOKENS=250
+
 # ===== CORS =====
 # 逗号分隔的允许跨域来源列表，生产环境绝不能使用 *
 CORS_ALLOW_ORIGINS=http://localhost:5173
+

.gitignore

@@ -62,3 +62,6 @@ logs/
 
 # codex
 .agents
+
+# personal local records (never commit)
+local/

QUICK_DEPLOY.md

@@ -390,12 +390,38 @@ Demo-glm/
 | 下载文档 | `/api/v1/documents/download/{doc_id}` | GET | 下载原文PDF/DOCX |
 | 文档列表 | `/api/v1/documents/list` | GET | 列出已上传文档 |
 | 检索知识 | `/api/v1/knowledge/search` | POST | 向量检索 |
-| 单次问答 | `/api/v1/agent/ask` | POST | 智能问答 |
-| 多轮对话 | `/api/v1/agent/chat` | POST | 会话对话 |
+| 单次问答 | `/api/v1/agent/ask` | POST | 标准单轮问答 |
+| 多轮对话 | `/api/v1/agent/chat` | POST | 标准会话对话 |
+| 流式对话 | `/api/v1/agent/chat/stream` | POST | 标准流式问答 (SSE) |
+| **Agentic 流式对话** | **`/api/v1/agent/agentic/stream`** | **POST** | **P0-1 多步推理 (SSE)：意图分析→查询分解→迭代检索→引文锚定→生成** |
 | 会话信息 | `/api/v1/agent/session/{id}` | GET | 获取会话 |
 | 删除会话 | `/api/v1/agent/session/{id}` | DELETE | 删除会话 |
-| Prompt模板 | `/api/v1/agent/templates` | GET | 模板列表 |
-| 可用模型 | `/api/v1/agent/models` | GET | LLM模型列表 |
+| 会话历史 | `/api/v1/agent/session/{id}/history` | GET | 获取历史记录 |
+| 会话列表 | `/api/v1/agent/sessions` | GET | 列出所有会话 |
+
+### Agentic 流式接口说明 (`/api/v1/agent/agentic/stream`)
+
+**请求体** (同 `/agent/chat/stream`)：
+```json
+{ "query": "GB 18384 与 ECE R100 在电池安全上有哪些差异？", "session_id": null, "top_k": 5 }
+```
+
+**额外 SSE 事件** (`thinking`)：
+```
+event: thinking
+data: {"step": "intent_analysis", "status": "done", "intent_type": "compare", "requires_decomposition": true}
+
+event: thinking
+data: {"step": "query_planning", "status": "done", "sub_queries": ["GB 18384 电池安全要求", "ECE R100 电池安全要求"]}
+
+event: thinking
+data: {"step": "retrieving", "status": "done", "query": "GB 18384 电池安全要求", "index": 1, "total": 2, "found": 8}
+
+event: thinking
+data: {"step": "grounding_check", "status": "done", "sufficient": true, "confidence": 0.82, "reason": "检索置信度充足"}
+```
+
+**意图类型**：`simple_qa`（单跳）/ `compare`（对比）/ `multi_hop`（多跳）/ `ambiguous`（模糊）
 
 ---
 

backend/app/api/models/agent.py

@@ -42,6 +42,11 @@ class ChatRequest(BaseModel):
     provider: Optional[str] = None
     model: Optional[str] = None
     top_k: Optional[int] = Field(default=None, ge=1, le=20)
+    # Optional document text uploaded by the user as conversation context.
+    # The text is injected directly into the LLM prompt so the model can
+    # answer questions about it without vector-store indexing.
+    context_text: Optional[str] = Field(default=None, max_length=12000)
+    context_filename: Optional[str] = Field(default=None, max_length=256)
 
 
 class ChatResponse(BaseModel):

backend/app/api/routes/agent.py

@@ -20,7 +20,11 @@ from app.api.models import (
 )
 from app.config.settings import settings
 from app.shared.async_utils import iter_in_thread
-from app.shared.bootstrap import get_agent_conversation_service, get_agent_session_service
+from app.shared.bootstrap import (
+    get_agent_conversation_service,
+    get_agent_session_service,
+    get_agentic_conversation_service,
+)
 # Keep route handlers close to their transport-layer wiring for easier auditing.
 
 
@@ -182,3 +186,58 @@ async def submit_feedback(request: FeedbackRequest):
         return {"message": "反馈已提交", "session_id": result.session_id, "message_index": result.message_index}
     except ValueError as exc:
         raise HTTPException(status_code=404, detail=str(exc))
+
+
+# ── P0-1: Agentic RAG endpoint ────────────────────────────────────────────────
+
+@router.post("/agentic/stream")
+async def agentic_stream(request: ChatRequest):
+    """Stream an Agentic RAG response with live multi-step reasoning trace.
+
+    Unlike the standard ``/chat/stream`` endpoint this route runs a full pipeline:
+    intent analysis → query planning → iterative retrieval → grounding check →
+    answer generation.
+
+    Extra SSE event types beyond the standard ones:
+
+    * ``thinking`` — reasoning sub-step progress; data is a JSON object with
+      ``step`` (intent_analysis / query_planning / retrieving / grounding_check),
+      ``status`` (running / done), and step-specific fields.
+
+    The ``sources``, ``content``, and ``done`` events are identical to the standard
+    chat-stream contract so the existing frontend parser can handle them without
+    changes.
+    """
+    async def generate_sse() -> AsyncGenerator[str, None]:
+        """Handle SSE generation for the agentic chat endpoint."""
+        try:
+            session_id_, event_stream = get_agentic_conversation_service().stream_agentic_chat(
+                query=request.query,
+                session_id=request.session_id,
+                filters=request.filters,
+                provider=request.provider or settings.llm_provider,
+                model=request.model or settings.llm_model,
+                top_k=request.top_k or settings.rag_top_k,
+                context_text=request.context_text,
+                context_filename=request.context_filename,
+            )
+            yield f"event: session\ndata: {json.dumps({'session_id': session_id_})}\n\n"
+            async for event_data in iter_in_thread(event_stream):
+                event_type = event_data.get("event", "content")
+                data = event_data.get("data", "")
+                if isinstance(data, (dict, list)):
+                    yield f"event: {event_type}\ndata: {json.dumps(data, ensure_ascii=False)}\n\n"
+                else:
+                    yield f"event: {event_type}\ndata: {data}\n\n"
+        except Exception as exc:
+            yield f"event: error\ndata: {str(exc)}\n\n"
+
+    return StreamingResponse(
+        generate_sse(),
+        media_type="text/event-stream",
+        headers={
+            "Cache-Control": "no-cache",
+            "Connection": "keep-alive",
+            "X-Accel-Buffering": "no",
+        },
+    )

backend/app/api/routes/compliance.py

@@ -85,9 +85,10 @@ async def analyze_stream(
     Events: stage | source | finding | done | error
     """
     from app.application.compliance.pipeline import (
+        detect_cross_clause_conflicts,
         extract_text_from_doc_id,
         extract_text_from_file,
-        run_clauses_parallel,
+        run_clauses_streaming,
         split_into_clauses,
         synthesize_conclusion,
     )
@@ -135,23 +136,27 @@ async def analyze_stream(
             await asyncio.sleep(0)
             clauses: list[str] = await asyncio.to_thread(split_into_clauses, para_text, client)
 
-            # ── Stage 3: retrieve + gap check (parallel across all clauses) ────────────
+            # ── Stage 3: progressive per-clause retrieve + gap check ──────
             findings: list[dict] = []
+            total_clauses = len(clauses)
 
             yield _sse({
                 "type": "stage",
                 "stage": "analyzing",
-                "label": f"Analyzing {len(clauses)} clauses in parallel…",
+                "label": f"Analyzing {total_clauses} clauses…",
             })
+            # Emit initial progress so the frontend can show the total count
+            yield _sse({"type": "progress", "done": 0, "total": total_clauses})
             await asyncio.sleep(0)
 
-            clause_results = await run_clauses_parallel(
+            done_count = 0
+            # Stream results as each clause completes (not after all finish)
+            async for res in run_clauses_streaming(
                 clauses, retrieval_service, client,
                 top_k=5,
                 domains=domains or None,
-            )
-
-            for res in clause_results:
+            ):
+                done_count += 1
                 i = res["index"]
                 chunks = res["chunks"]
                 finding = res["finding"]
@@ -165,14 +170,25 @@ async def analyze_stream(
                         "score": round(float(getattr(chunk, "score", 0)), 3),
                         "status": "retrieved",
                         "full_content": (getattr(chunk, "text", "") or "")[:300],
+                        "clause_index": i,
                     })
 
                 if finding:
                     findings.append(finding)
                     yield _sse({"type": "finding", **finding})
 
+                # Real progress update after each clause completes
+                yield _sse({"type": "progress", "done": done_count, "total": total_clauses})
                 await asyncio.sleep(0)
 
+            # ── Stage 3b: cross-clause conflict detection ─────────────────
+            if findings:
+                conflicts = await asyncio.to_thread(
+                    detect_cross_clause_conflicts, findings, client
+                )
+                if conflicts:
+                    yield _sse({"type": "conflicts", "items": conflicts})
+
             # ── Stage 4: synthesize conclusion ────────────────────────────
             yield _sse({"type": "stage", "stage": "concluding", "label": "Generating conclusion…"})
             await asyncio.sleep(0)

backend/app/api/routes/documents.py

@@ -241,6 +241,9 @@ async def get_document_management_list():
                 "updated_at": item.updated_at.isoformat(),
                 "regulation_type": item.regulation_type,
                 "version": item.version,
+                # True only when the original binary file is stored in MinIO.
+                # Milvus-only synthetic docs have no binary file — download is disabled.
+                "has_file": bool(item.object_name),
             }
             for item in documents
         ],

backend/app/api/routes/rag.py

@@ -3,10 +3,14 @@
 from __future__ import annotations
 
 import json
-from typing import AsyncGenerator
+import os
+import re
+import tempfile
+from typing import AsyncGenerator, Optional
 
-from fastapi import APIRouter, Depends
+from fastapi import APIRouter, Depends, File, UploadFile
 from fastapi.responses import StreamingResponse
+from loguru import logger
 
 from app.api.dependencies.auth import get_current_user
 from app.config.settings import settings
@@ -15,6 +19,8 @@ from app.schemas.rag import RagChatRequest, QuickQuestionsResponse, QuickQuestio
 from app.shared.async_utils import iter_in_thread
 from app.shared.bootstrap import get_agent_conversation_service
 
+# Maximum characters of document text injected as LLM context (≈ 6 000 tokens).
+_MAX_CONTEXT_CHARS = 8_000
 
 router = APIRouter(prefix="/rag", tags=["RAG问答"])
 
@@ -28,17 +34,90 @@ _DEFAULT_QUICK_QUESTIONS = [
 ]
 
 
+def _extract_text_from_bytes(content: bytes, filename: str) -> str:
+    """Extract plain text from an uploaded file using the document parser.
+
+    Tries the configured parser first; falls back to raw UTF-8 decode for
+    plain-text formats (.txt, .md). Returns at most _MAX_CONTEXT_CHARS characters
+    so the text fits comfortably inside the LLM context window.
+    """
+    suffix = os.path.splitext(filename or "doc.pdf")[1] or ".pdf"
+    # Fast path: plain-text files don't need a parser
+    if suffix.lower() in {".txt", ".md", ".csv"}:
+        try:
+            return content.decode("utf-8", errors="replace")[:_MAX_CONTEXT_CHARS]
+        except Exception:
+            pass
+
+    tmp_path = ""
+    try:
+        with tempfile.NamedTemporaryFile(delete=False, suffix=suffix) as tmp:
+            tmp.write(content)
+            tmp_path = tmp.name
+        from app.shared.bootstrap import get_document_command_service
+        svc = get_document_command_service()
+        parsed = svc.parser.parse(file_path=tmp_path, doc_id="ctx_extract", doc_name=filename)
+        if parsed.raw_text:
+            return parsed.raw_text[:_MAX_CONTEXT_CHARS]
+        # Fallback: join semantic blocks
+        return "\n".join(
+            b.get("text", "") for b in parsed.semantic_blocks if b.get("text")
+        )[:_MAX_CONTEXT_CHARS]
+    except Exception as exc:
+        logger.warning("Context text extraction failed for {}: {}", filename, exc)
+        return ""
+    finally:
+        if tmp_path:
+            try:
+                os.unlink(tmp_path)
+            except OSError:
+                pass
+
+
+@router.post("/upload-context")
+async def upload_context(
+    file: UploadFile = File(...),
+    current_user: UserClaims = Depends(get_current_user),
+):
+    """Extract text from an uploaded document and return it as conversation context.
+
+    The client stores the returned text and includes it in subsequent /rag/chat
+    requests via the context_text field — the LLM receives the document content
+    directly without requiring vector-store indexing.
+    """
+    content = await file.read()
+    filename = file.filename or "document"
+    text = await __import__("asyncio").to_thread(_extract_text_from_bytes, content, filename)
+    if not text.strip():
+        from fastapi import HTTPException
+        raise HTTPException(status_code=422, detail="Could not extract text from the uploaded file.")
+    return {
+        "filename": filename,
+        "text": text,
+        "char_count": len(text),
+        "truncated": len(text) >= _MAX_CONTEXT_CHARS,
+    }
+
+
 @router.post("/chat")
 async def rag_chat(
     request: RagChatRequest,
     current_user: UserClaims = Depends(get_current_user),
 ):
-    """Stream RAG Q&A using the real agent service."""
+    """Stream RAG Q&A using the real agent service.
+
+    When request.context_text is provided the document text is passed directly
+    to the answer generator as a dedicated document context section — RAG
+    retrieval still runs on the user's original question (not the document text)
+    so embedding quality is preserved for regulation chunk matching.
+    """
     session_id, event_stream = get_agent_conversation_service().stream_chat(
         query=request.query,
         session_id=request.session_id,
         filters=request.filters,
         top_k=request.top_k or settings.rag_top_k,
+        context_text=request.context_text,
+        context_filename=request.context_filename,
     )
 
     async def generate() -> AsyncGenerator[str, None]:

backend/app/application/agent/__init__.py

@@ -1,7 +1,13 @@
 """Initialize the app.application.agent package."""
 
 from .services import AgentConversationService, AgentSessionFeedbackResult, AgentSessionService
+from .agentic_service import AgenticConversationService
 # Keep package boundaries explicit so backend imports stay predictable.
 
 
-__all__ = ["AgentConversationService", "AgentSessionFeedbackResult", "AgentSessionService"]
+__all__ = [
+    "AgentConversationService",
+    "AgentSessionFeedbackResult",
+    "AgentSessionService",
+    "AgenticConversationService",
+]

backend/app/application/agent/agentic_service.py

@@ -0,0 +1,453 @@
+"""Implement the Agentic RAG pipeline for multi-step reasoning (P0-1).
+
+Architecture
+------------
+The pipeline adds four explicit reasoning steps before answer generation:
+
+  1. Intent Analysis  — classify query type (simple_qa / compare / multi_hop / ambiguous)
+  2. Query Planning   — for complex intents, decompose into focused sub-queries
+  3. Iterative Retrieval — retrieve for each sub-query, merge with deduplication
+  4. Grounding Check  — verify retrieved context is sufficient; refine query when not
+  5. Answer Generation — stream final answer with citations (reuses AnswerGenerator)
+
+Each step emits SSE ``thinking`` events so the frontend can render the live
+reasoning trace.  The pipeline is entirely synchronous and returns a generator so
+it plugs into the same ``iter_in_thread`` pattern used by the existing chat routes.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field
+from typing import Generator
+
+from loguru import logger
+
+from app.application.knowledge import KnowledgeRetrievalService
+from app.application.agent.hyde_expander import HyDEExpander
+from app.config.settings import settings
+from app.domain.conversation import ConversationStore
+from app.domain.retrieval import RetrievedChunk
+from app.infrastructure.llm.openai_compatible_answer_generator import OpenAICompatibleAnswerGenerator
+from app.services.llm.llm_factory import get_llm_client
+
+# ── Prompts ───────────────────────────────────────────────────────────────────
+# Each prompt is kept module-level for easy review and fine-tuning.
+
+_INTENT_SYSTEM = (
+    "You are a query classifier for a Chinese regulatory compliance knowledge base.\n\n"
+    "Classify the query into exactly one of:\n"
+    '- "simple_qa"   : Single-hop, factual question about one regulation or clause\n'
+    '- "compare"     : Comparison between two or more regulations, standards, or versions\n'
+    '- "multi_hop"   : Requires chaining facts across multiple regulations to answer\n'
+    '- "ambiguous"   : Too vague or broad to retrieve effectively\n\n'
+    "Return ONLY valid JSON — no markdown, no extra text:\n"
+    '{"type": "...", "reason": "one sentence", "requires_decomposition": true/false}\n\n'
+    '"requires_decomposition" must be true for compare and multi_hop types.'
+)
+
+_PLAN_SYSTEM = (
+    "You are a query planner for a Chinese regulatory compliance knowledge base.\n\n"
+    "Decompose the query into 2-4 focused, self-contained sub-queries that together fully "
+    "address the original question.  Each sub-query must target one specific regulation, "
+    "clause, or concept and be independently searchable.\n\n"
+    "Return ONLY a valid JSON array — no markdown, no extra text:\n"
+    '["sub-query 1", "sub-query 2", ...]'
+)
+
+_GROUNDING_SYSTEM = (
+    "You are a grounding verifier for a regulatory compliance QA system.\n\n"
+    "Given a query and retrieved regulation passages, decide whether the passages contain "
+    "sufficient, accurate information to answer the query.\n\n"
+    "Return ONLY valid JSON — no markdown, no extra text:\n"
+    '{"sufficient": true/false, "confidence": 0.0-1.0, "reason": "one sentence", '
+    '"refined_query": "a more specific search query if not sufficient, else null"}'
+)
+
+
+# ── Result dataclasses ────────────────────────────────────────────────────────
+
+@dataclass
+class IntentResult:
+    """Capture the output of the intent-analysis step."""
+
+    type: str = "simple_qa"
+    reason: str = ""
+    requires_decomposition: bool = False
+
+
+@dataclass
+class GroundingResult:
+    """Capture the output of the grounding-check step."""
+
+    sufficient: bool = True
+    confidence: float = 1.0
+    reason: str = ""
+    refined_query: str | None = None
+
+
+# ── Service ───────────────────────────────────────────────────────────────────
+
+class AgenticConversationService:
+    """Multi-step Agentic RAG pipeline with live reasoning trace via SSE.
+
+    The service is intentionally synchronous so it can be wrapped in
+    ``iter_in_thread`` by the route layer without any async boilerplate.
+    """
+
+    def __init__(
+        self,
+        *,
+        retrieval_service: KnowledgeRetrievalService,
+        answer_generator: OpenAICompatibleAnswerGenerator,
+        conversation_store: ConversationStore,
+    ) -> None:
+        """Initialise with injected dependencies from the composition root."""
+        self.retrieval_service = retrieval_service
+        self.answer_generator = answer_generator
+        self.conversation_store = conversation_store
+        # HyDE expander is stateless — one instance shared for all requests.
+        self._hyde = HyDEExpander()
+
+    # ── Private helpers ───────────────────────────────────────────────────────
+
+    def _llm_json(
+        self,
+        system: str,
+        user: str,
+        provider: str | None,
+        model: str | None,
+        max_tokens: int = 300,
+    ) -> dict | list | None:
+        """Call the LLM with a JSON-only prompt and return the parsed result.
+
+        Returns ``None`` on any API or parse failure so callers can degrade
+        gracefully without raising.
+        """
+        client = get_llm_client(
+            provider=provider or settings.llm_provider,
+            model=model or settings.llm_model,
+        )
+        resp = client.chat(
+            [{"role": "system", "content": system}, {"role": "user", "content": user}],
+            max_tokens=max_tokens,
+            temperature=0.1,
+        )
+        if not resp.is_success:
+            logger.warning("AgenticService LLM call failed: {}", resp.error)
+            return None
+        try:
+            raw = resp.content.strip()
+            # Strip accidental markdown code fences the model may add.
+            if raw.startswith("```"):
+                parts = raw.split("```")
+                raw = parts[1] if len(parts) > 1 else raw
+                if raw.startswith("json"):
+                    raw = raw[4:]
+            return json.loads(raw.strip())
+        except (json.JSONDecodeError, IndexError) as exc:
+            logger.debug("AgenticService JSON parse failed: {} | raw={}", exc, resp.content[:200])
+            return None
+
+    def _analyze_intent(
+        self, query: str, provider: str | None, model: str | None
+    ) -> IntentResult:
+        """Classify query intent to select the appropriate retrieval strategy."""
+        data = self._llm_json(
+            _INTENT_SYSTEM,
+            f"Query: {query}",
+            provider,
+            model,
+            max_tokens=settings.agentic_intent_max_tokens,
+        )
+        if isinstance(data, dict):
+            return IntentResult(
+                type=str(data.get("type", "simple_qa")),
+                reason=str(data.get("reason", "")),
+                requires_decomposition=bool(data.get("requires_decomposition", False)),
+            )
+        return IntentResult(type="simple_qa", reason="fallback — classifier returned no JSON", requires_decomposition=False)
+
+    def _plan_queries(
+        self, query: str, intent_type: str, provider: str | None, model: str | None
+    ) -> list[str]:
+        """Decompose a complex query into focused, independently-retrievable sub-queries."""
+        data = self._llm_json(
+            _PLAN_SYSTEM,
+            f"Original query ({intent_type}): {query}",
+            provider,
+            model,
+            max_tokens=settings.agentic_plan_max_tokens,
+        )
+        if isinstance(data, list) and data:
+            # Cap at configured maximum to keep latency predictable.
+            return [str(q) for q in data[:settings.agentic_max_sub_queries] if q]
+        return [query]
+
+    def _check_grounding(
+        self,
+        query: str,
+        chunks: list[RetrievedChunk],
+        provider: str | None,
+        model: str | None,
+    ) -> GroundingResult:
+        """Verify whether retrieved chunks are sufficient to ground an accurate answer.
+
+        Uses a fast score-threshold heuristic first; falls back to an LLM call only
+        when scores are borderline so that the happy-path adds no extra latency.
+        """
+        if not chunks:
+            return GroundingResult(
+                sufficient=False,
+                confidence=0.0,
+                reason="未检索到相关内容",
+                refined_query=None,
+            )
+
+        avg_score = sum(c.score for c in chunks) / len(chunks)
+        # Fast path: high-confidence retrieval → skip extra LLM call.
+        if avg_score > settings.agentic_grounding_threshold and len(chunks) >= 3:
+            return GroundingResult(
+                sufficient=True,
+                confidence=round(avg_score, 3),
+                reason="检索置信度充足，无需二次查询",
+                refined_query=None,
+            )
+
+        # LLM-based grounding check for borderline retrievals.
+        context_preview = "\n".join(
+            f"[{i + 1}] (score={c.score:.2f}) {c.text[:200]}" for i, c in enumerate(chunks[:5])
+        )
+        data = self._llm_json(
+            _GROUNDING_SYSTEM,
+            f"Query: {query}\n\nRetrieved passages:\n{context_preview}",
+            provider,
+            model,
+            max_tokens=settings.agentic_grounding_max_tokens,
+        )
+        if isinstance(data, dict):
+            return GroundingResult(
+                sufficient=bool(data.get("sufficient", True)),
+                confidence=float(data.get("confidence", 0.5)),
+                reason=str(data.get("reason", "")),
+                refined_query=data.get("refined_query") or None,
+            )
+        return GroundingResult(sufficient=True, confidence=0.5, reason="grounding check skipped (parse error)", refined_query=None)
+
+    @staticmethod
+    def _intent_to_template(intent_type: str) -> str:
+        """Map an intent type to the best prompt template name for answer generation."""
+        mapping = {
+            "compare": "comparison",
+            "multi_hop": "compliance_qa",
+            "simple_qa": "compliance_qa",
+            "ambiguous": "compliance_qa",
+        }
+        return mapping.get(intent_type, "compliance_qa")
+
+    @staticmethod
+    def _deduplicate(chunks: list[RetrievedChunk], max_chunks: int) -> list[RetrievedChunk]:
+        """Remove duplicate chunk IDs, preserving first-occurrence order up to max_chunks."""
+        seen: set[str] = set()
+        result: list[RetrievedChunk] = []
+        for chunk in chunks:
+            if chunk.chunk_id not in seen:
+                seen.add(chunk.chunk_id)
+                result.append(chunk)
+                if len(result) >= max_chunks:
+                    break
+        return result
+
+    # ── Public interface ──────────────────────────────────────────────────────
+
+    def stream_agentic_chat(
+        self,
+        *,
+        query: str,
+        session_id: str | None = None,
+        filters: str | None = None,
+        provider: str | None = None,
+        model: str | None = None,
+        top_k: int = 5,
+        context_text: str | None = None,
+        context_filename: str | None = None,
+    ) -> tuple[str, Generator[dict, None, None]]:
+        """Run the full Agentic RAG pipeline and return ``(session_id, event_generator)``.
+
+        When context_text is provided (user-attached document) it is:
+        - Summarised and prepended to the intent-analysis prompt so the classifier
+          understands what kind of question is being asked.
+        - Treated as baseline grounding so the pipeline skips unnecessary retries
+          when the document itself is the primary source.
+        - Passed to the answer generator so the LLM sees the full document alongside
+          retrieved regulation chunks.
+
+        The generator yields SSE event dicts compatible with the route's
+        ``iter_in_thread`` pattern.
+        """
+        session = self.conversation_store.get_session(session_id) if session_id else None
+        if session is None:
+            session = self.conversation_store.create_session()
+        self.conversation_store.save_message(session.session_id, role="user", content=query)
+        history = [{"role": msg.role, "content": msg.content} for msg in session.messages[-10:]]
+        active_session_id = session.session_id
+
+        # Build a brief document summary for classifier/planner prompts (avoid
+        # passing the full text which could overwhelm small-context LLMs).
+        _doc_summary: str = ""
+        if context_text and context_text.strip():
+            _doc_label = context_filename or "document"
+            _preview = context_text.strip()[:400]
+            _doc_summary = f"[User has attached document: {_doc_label}]\nDocument preview: {_preview}…\n\n"
+
+        def event_stream() -> Generator[dict, None, None]:
+            """Execute all pipeline steps and yield SSE events."""
+            # ── Step 1: Intent Analysis ──────────────────────────────────────
+            yield {"event": "thinking", "data": {"step": "intent_analysis", "status": "running"}}
+            # Prepend doc summary so the classifier knows what the user is asking about
+            intent_user_msg = f"{_doc_summary}Query: {query}" if _doc_summary else f"Query: {query}"
+            data = self._llm_json(
+                _INTENT_SYSTEM, intent_user_msg, provider, model,
+                max_tokens=settings.agentic_intent_max_tokens,
+            )
+            if isinstance(data, dict):
+                intent = IntentResult(
+                    type=str(data.get("type", "simple_qa")),
+                    reason=str(data.get("reason", "")),
+                    requires_decomposition=bool(data.get("requires_decomposition", False)),
+                )
+            else:
+                intent = IntentResult(type="simple_qa", reason="fallback", requires_decomposition=False)
+            logger.debug("Agentic intent: type={} decompose={}", intent.type, intent.requires_decomposition)
+            yield {
+                "event": "thinking",
+                "data": {
+                    "step": "intent_analysis",
+                    "status": "done",
+                    "intent_type": intent.type,
+                    "reason": intent.reason,
+                    "requires_decomposition": intent.requires_decomposition,
+                },
+            }
+
+            # ── Step 2: Query Planning ───────────────────────────────────────
+            sub_queries: list[str] = [query]
+            if intent.requires_decomposition:
+                yield {"event": "thinking", "data": {"step": "query_planning", "status": "running"}}
+                plan_user_msg = f"{_doc_summary}Original query ({intent.type}): {query}" if _doc_summary else f"Original query ({intent.type}): {query}"
+                data_plan = self._llm_json(
+                    _PLAN_SYSTEM, plan_user_msg, provider, model,
+                    max_tokens=settings.agentic_plan_max_tokens,
+                )
+                if isinstance(data_plan, list) and data_plan:
+                    sub_queries = [str(q) for q in data_plan[:settings.agentic_max_sub_queries] if q]
+                logger.debug("Agentic sub-queries ({}): {}", len(sub_queries), sub_queries)
+                yield {
+                    "event": "thinking",
+                    "data": {"step": "query_planning", "status": "done", "sub_queries": sub_queries},
+                }
+
+            # ── Step 3: Iterative Retrieval ──────────────────────────────────
+            # Always retrieve using the user's original question (NOT the document
+            # text) so embedding quality is preserved for regulation matching.
+            # HyDE enriches the retrieval query with a short hypothetical answer
+            # to close the vocabulary gap between terse queries and long documents.
+            candidate_k = max(top_k * 3, 15)
+            all_chunks: list[RetrievedChunk] = []
+
+            # For simple_qa with a single query, HyDE gives the biggest benefit
+            # (bridging vague/colloquial questions to formal document language).
+            # For compare/multi_hop, the planner already decomposed into precise
+            # sub-queries, so HyDE is less critical but still applied per sub-query.
+            for idx, sq in enumerate(sub_queries, start=1):
+                yield {
+                    "event": "thinking",
+                    "data": {"step": "retrieving", "status": "running", "query": sq, "index": idx, "total": len(sub_queries)},
+                }
+                # HyDE expansion: generate hypothetical answer, embed it for retrieval.
+                # Falls back to original sub-query if LLM call fails.
+                retrieval_query = self._hyde.expand(sq)
+                chunks = self.retrieval_service.retrieve(query=retrieval_query, top_k=candidate_k, filters=filters)
+                all_chunks.extend(chunks)
+                yield {
+                    "event": "thinking",
+                    "data": {"step": "retrieving", "status": "done", "query": sq, "index": idx, "total": len(sub_queries), "found": len(chunks)},
+                }
+
+            unique_chunks = self._deduplicate(all_chunks, max_chunks=top_k * 4)
+
+            # ── Step 4: Grounding Check ──────────────────────────────────────
+            yield {"event": "thinking", "data": {"step": "grounding_check", "status": "running"}}
+
+            # When the user has attached a document, the document itself provides
+            # baseline grounding — skip the re-query loop to avoid the LLM asking
+            # "please provide the document text" as a refined query.
+            if context_text and context_text.strip():
+                grounding = GroundingResult(
+                    sufficient=True,
+                    confidence=0.95,
+                    reason="用户已附件上传文档，以文档内容为基础作答",
+                    refined_query=None,
+                )
+            else:
+                grounding = self._check_grounding(query, unique_chunks, provider, model)
+
+            yield {
+                "event": "thinking",
+                "data": {
+                    "step": "grounding_check",
+                    "status": "done",
+                    "sufficient": grounding.sufficient,
+                    "confidence": grounding.confidence,
+                    "reason": grounding.reason,
+                },
+            }
+
+            # Only retry from vector store when no document is attached and grounding failed
+            if not grounding.sufficient and grounding.refined_query and not context_text:
+                logger.info("Grounding insufficient — re-querying: {}", grounding.refined_query)
+                yield {
+                    "event": "thinking",
+                    "data": {"step": "retrieving", "status": "running", "query": grounding.refined_query, "index": 1, "total": 1, "retry": True},
+                }
+                # Apply HyDE to the refined query as well for better retrieval.
+                refined_hyde_query = self._hyde.expand(grounding.refined_query)
+                refined_chunks = self.retrieval_service.retrieve(query=refined_hyde_query, top_k=candidate_k, filters=filters)
+                all_chunks.extend(refined_chunks)
+                unique_chunks = self._deduplicate(all_chunks, max_chunks=top_k * 4)
+                yield {
+                    "event": "thinking",
+                    "data": {"step": "retrieving", "status": "done", "query": grounding.refined_query, "index": 1, "total": 1, "found": len(refined_chunks), "retry": True},
+                }
+
+            final_chunks = unique_chunks[:top_k]
+
+            # ── Step 5: Answer Generation ────────────────────────────────────
+            sources_payload = [s.__dict__ for s in self.answer_generator._sources(final_chunks)]
+            yield {"event": "sources", "data": sources_payload}
+
+            answer_parts: list[str] = []
+            for event in self.answer_generator.stream_generate(
+                query=query,
+                retrieved_chunks=final_chunks,
+                history=history,
+                provider=provider,
+                model=model,
+                prompt_template=self._intent_to_template(intent.type),
+                context_text=context_text,
+                context_filename=context_filename,
+            ):
+                if event.get("event") == "content":
+                    answer_parts.append(str(event.get("data", "")))
+                yield event
+
+            full_answer = "".join(answer_parts)
+            self.conversation_store.save_message(
+                active_session_id,
+                role="assistant",
+                content=full_answer,
+                sources=sources_payload,
+            )
+
+        return active_session_id, event_stream()
+

backend/app/application/agent/hyde_expander.py

@@ -0,0 +1,105 @@
+"""Implement HyDE (Hypothetical Document Embeddings) query expansion.
+
+HyDE improves dense retrieval by addressing the vocabulary gap between
+short user queries and longer document passages:
+
+  User query  →  [LLM generates hypothetical answer]
+                         ↓
+             embed hypothetical answer (not original query)
+                         ↓
+             retrieve similar real passages from Milvus
+
+The hypothetical answer uses the same vocabulary and phrasing as documents,
+so its embedding is much closer to relevant chunks than a terse query embedding.
+
+Usage:
+    expander = HyDEExpander()
+    retrieval_query = expander.expand(query, provider=..., model=...)
+    chunks = retrieval_service.retrieve(query=retrieval_query, ...)
+
+When the LLM call fails, expand() falls back to the original query so the
+retrieval pipeline degrades gracefully.
+
+References:
+    Gao et al. (2022), "Precise Zero-Shot Dense Retrieval without Relevance Labels"
+    https://arxiv.org/abs/2212.10496
+"""
+
+from __future__ import annotations
+
+from loguru import logger
+
+from app.config.settings import settings
+from app.services.llm.llm_factory import get_llm_client
+
+# Maximum chars to trim from the hypothetical answer to avoid token overrun.
+_MAX_HYPOTHESIS_CHARS = 600
+
+# System prompt that instructs the LLM to write a passage *as if* it were
+# from a regulatory document, not a conversation answer.
+_HYDE_SYSTEM = (
+    "你是一位法规知识库专家。用户提出了一个问题，"
+    "请用50-120字写一段话，模拟如果相关法规文档中存在完美答案，"
+    "该段落会是什么内容。\n\n"
+    "要求：\n"
+    "- 使用与法规文档相同的正式书面语气\n"
+    "- 包含可能的条款编号、标准名称等关键术语\n"
+    "- 不要解释你在做什么，直接输出假设性段落\n"
+    "- 如问题过于模糊，写一段合理的通用法规说明"
+)
+
+
+class HyDEExpander:
+    """Generate a hypothetical document passage to improve dense retrieval.
+
+    The expander is stateless — instantiate once and call expand() per query.
+    It requires no external dependencies beyond the project's existing LLM
+    client infrastructure.
+    """
+
+    def expand(self, query: str) -> str:
+        """Return a combined retrieval query: original query + hypothetical passage.
+
+        The combination ensures:
+        - Dense retrieval uses the enriched hypothetical text (semantic match).
+        - BM25 retrieval still benefits from the original query keywords.
+
+        The model used is ``settings.hyde_llm_model`` (dedicated lightweight model)
+        falling back to the main ``settings.llm_model`` when not configured.
+
+        If the LLM call fails for any reason, returns the original query unchanged.
+        """
+        if not settings.hyde_enabled:
+            return query
+
+        # Use the dedicated HyDE model when configured; fall back to main LLM.
+        # A lightweight model (e.g. qwen3.5-flash) is sufficient for generating
+        # a short hypothetical passage and significantly reduces cost + latency.
+        provider = settings.hyde_llm_provider or settings.llm_provider
+        model = settings.hyde_llm_model or settings.llm_model
+
+        try:
+            client = get_llm_client(provider=provider, model=model)
+            resp = client.chat(
+                messages=[
+                    {"role": "system", "content": _HYDE_SYSTEM},
+                    {"role": "user", "content": f"问题：{query}"},
+                ],
+                max_tokens=settings.hyde_max_tokens,
+                # Low temperature: we want a plausible, deterministic passage.
+                temperature=0.3,
+            )
+            if not resp.is_success or not resp.content:
+                logger.debug("HyDE LLM call failed or empty — using original query")
+                return query
+
+            hypothesis = resp.content.strip()[:_MAX_HYPOTHESIS_CHARS]
+            logger.debug("HyDE expanded query ({}→{} chars)", len(query), len(hypothesis))
+
+            # Concatenate: the embedding model will see the full combined text,
+            # so the resulting vector leans toward the hypothetical document style.
+            return f"{query}\n\n{hypothesis}"
+
+        except Exception as exc:  # noqa: BLE001 — intentional broad catch for graceful fallback
+            logger.warning("HyDE expansion failed: {} — using original query", exc)
+            return query

backend/app/application/agent/services.py

@@ -9,6 +9,7 @@ from app.domain.conversation import AnswerGenerator, AnswerResult, ConversationS
 from app.domain.retrieval import RetrievedChunk
 
 from app.application.knowledge import KnowledgeRetrievalService
+from app.application.agent.hyde_expander import HyDEExpander
 # Keep orchestration logic centralized so use-case flow stays easy to trace.
 
 
@@ -26,6 +27,8 @@ class AgentConversationService:
         self.retrieval_service = retrieval_service
         self.answer_generator = answer_generator
         self.conversation_store = conversation_store
+        # Shared HyDE expander — stateless, safe for reuse across requests.
+        self._hyde = HyDEExpander()
 
     def ask(
         self,
@@ -108,14 +111,26 @@ class AgentConversationService:
         model: str | None = None,
         top_k: int = 5,
         prompt_template: str | None = None,
+        context_text: str | None = None,
+        context_filename: str | None = None,
     ) -> tuple[str, Generator[dict, None, None]]:
-        """Stream chat for the Agent Conversation Service instance."""
+        """Stream chat for the Agent Conversation Service instance.
+
+        When context_text is provided the user's document is passed directly to
+        the answer generator — RAG retrieval still runs on the user's question
+        (not the document text) to find relevant regulation passages.
+        """
         session = self.conversation_store.get_session(session_id) if session_id else None
         if session is None:
             session = self.conversation_store.create_session()
         self.conversation_store.save_message(session.session_id, role="user", content=query)
         history = [{"role": msg.role, "content": msg.content} for msg in session.messages[-10:]]
-        retrieved = self.retrieval_service.retrieve(query=query, top_k=top_k, filters=filters)
+        # HyDE: expand the query with a hypothetical answer to improve dense retrieval.
+        # For document-context queries, skip HyDE since the document itself guides retrieval.
+        retrieval_query = self._hyde.expand(query) if not context_text else query
+        # Retrieve using the enriched query — NOT the document text —
+        # so embedding quality is preserved for regulation chunk matching.
+        retrieved = self.retrieval_service.retrieve(query=retrieval_query, top_k=top_k, filters=filters)
 
         def event_stream() -> Generator[dict, None, None]:
             """Handle event stream for the Agent Conversation Service instance."""
@@ -129,6 +144,8 @@ class AgentConversationService:
                 provider=provider,
                 model=model,
                 prompt_template=prompt_template,
+                context_text=context_text,
+                context_filename=context_filename,
             ):
                 if event.get("event") == "sources":
                     sources_payload = event.get("data", [])
@@ -189,3 +206,4 @@ class AgentSessionService:
             raise ValueError("消息索引不存在")
         # Preserve the existing API behavior until a persistent feedback store is introduced.
         return AgentSessionFeedbackResult(session_id=session_id, message_index=message_index)
+

backend/app/application/compliance/pipeline.py

@@ -51,19 +51,36 @@ def _extract_json(text: str):
 
 
 def extract_text_from_doc_id(doc_id: str) -> str:
+    """Fetch the full text of a document by retrieving its chunks filtered by doc_id.
+
+    Uses a high top_k and doc_id filter to reconstruct the document in chunk order,
+    avoiding the previous approach of semantic search by doc_name which could return
+    chunks from unrelated documents.
+    """
     from app.shared.bootstrap import get_document_query_service, get_retrieval_service
     doc = get_document_query_service().get(doc_id)
     if not doc:
         raise ValueError(f"Document '{doc_id}' not found")
     service = get_retrieval_service()
-    chunks = service.retrieve(query=doc.doc_name, top_k=30)
-    doc_chunks = [c for c in chunks if c.doc_id == doc_id]
+    # Use doc_name as a broad query, filter strictly by doc_id so we only get
+    # this document's chunks; top_k=100 covers most real-world documents.
+    chunks = service.retrieve(query=doc.doc_name, top_k=100, filters=doc_id)
+    doc_chunks = [c for c in chunks if getattr(c, "doc_id", None) == doc_id]
     if not doc_chunks:
-        doc_chunks = chunks[:15]
-    return "\n\n".join(c.text for c in doc_chunks[:15])
+        # Fallback: use top results even without doc_id match (e.g., legacy store)
+        doc_chunks = chunks[:30]
+    # Sort by chunk_index to preserve document reading order
+    doc_chunks.sort(key=lambda c: getattr(c, "chunk_index", 0))
+    return "\n\n".join(c.text for c in doc_chunks[:40])
 
 
 def extract_text_from_file(content: bytes, filename: str) -> str:
+    """Parse an uploaded file and return its full text content.
+
+    Removed previous 4000-char cap so large specifications and standards are
+    fully analysed. The caller is responsible for splitting the text into
+    clause-sized chunks before passing to the LLM.
+    """
     from app.shared.bootstrap import get_document_command_service
     suffix = os.path.splitext(filename or "doc.pdf")[1] or ".pdf"
     tmp_path = ""
@@ -74,10 +91,11 @@ def extract_text_from_file(content: bytes, filename: str) -> str:
         service = get_document_command_service()
         parsed = service.parser.parse(file_path=tmp_path, doc_id="tmp_analysis", doc_name=filename)
         if parsed.raw_text:
-            return parsed.raw_text[:4000]
+            # Return full text — truncation happens in split_into_clauses()
+            return parsed.raw_text
         return "\n".join(
-            b.get("text", "") for b in parsed.semantic_blocks[:30] if b.get("text")
-        )[:4000]
+            b.get("text", "") for b in parsed.semantic_blocks if b.get("text")
+        )
     except Exception as exc:
         logger.warning("File text extraction failed: {}", exc)
         return ""
@@ -88,27 +106,68 @@ def extract_text_from_file(content: bytes, filename: str) -> str:
 
 
 def split_into_clauses(text: str, client: "BaseLLMClient") -> list[str]:
-    prompt = (
-        "You are a compliance analysis expert. Split the following text into 3-8 "
-        "semantically complete compliance clauses. Each clause should be an independent "
-        "compliance requirement or technical statement.\n"
-        "Return as JSON array of strings, e.g.:\n"
-        '["Clause one...", "Clause two..."]\n'
-        "Return ONLY the JSON array.\n\n"
-        f"Text:\n{text[:2000]}"
-    )
-    response = client.chat([{"role": "user", "content": prompt}], max_tokens=1000)
-    if response.is_success:
-        try:
-            result = _extract_json(response.content)
-            if isinstance(result, list):
-                clauses = [str(c).strip() for c in result if str(c).strip()]
-                if clauses:
-                    return clauses[:8]
-        except (ValueError, TypeError):
-            logger.warning("Clause split JSON parse failed, using fallback")
-    sentences = re.split(r"[.?!;\n]+", text)
-    return [s.strip() for s in sentences if len(s.strip()) > 20][:6]
+    """Split a compliance document into semantically independent clauses.
+
+    For long texts (> 2 000 chars) the document is processed in overlapping
+    2 000-char windows so no content is missed. Each window produces up to 4
+    clauses; results are deduplicated and capped at 12 total to keep analysis
+    latency reasonable.
+    """
+    # Window size and step for sliding-window clause extraction
+    _WINDOW = 2000
+    _STEP = 1800  # 200-char overlap to avoid cutting clauses at boundaries
+    _MAX_CLAUSES = 12
+
+    windows = []
+    if len(text) <= _WINDOW:
+        windows = [text]
+    else:
+        pos = 0
+        while pos < len(text):
+            windows.append(text[pos: pos + _WINDOW])
+            pos += _STEP
+
+    all_clauses: list[str] = []
+    for window in windows:
+        prompt = (
+            "You are a compliance analysis expert. Split the following text into "
+            "3-4 semantically complete compliance clauses. Each clause must be an "
+            "independent requirement or technical statement. Omit section headings, "
+            "definitions, and non-normative text.\n"
+            "Return as JSON array of strings, e.g.:\n"
+            '["Clause one...", "Clause two..."]\n'
+            "Return ONLY the JSON array.\n\n"
+            f"Text:\n{window}"
+        )
+        response = client.chat([{"role": "user", "content": prompt}], max_tokens=800)
+        if response.is_success:
+            try:
+                result = _extract_json(response.content)
+                if isinstance(result, list):
+                    clauses = [str(c).strip() for c in result if str(c).strip()]
+                    all_clauses.extend(clauses[:4])
+            except (ValueError, TypeError):
+                logger.warning("Clause split JSON parse failed for window, using sentence fallback")
+                sentences = re.split(r"[.?!;\n]+", window)
+                all_clauses.extend(s.strip() for s in sentences if len(s.strip()) > 20)
+        else:
+            # LLM unavailable — fall back to sentence splitting for this window
+            sentences = re.split(r"[.?!;\n]+", window)
+            all_clauses.extend(s.strip() for s in sentences if len(s.strip()) > 20)
+
+        if len(all_clauses) >= _MAX_CLAUSES:
+            break
+
+    # Deduplicate near-duplicates (same first 80 chars) that span window boundaries
+    seen: set[str] = set()
+    deduped: list[str] = []
+    for c in all_clauses:
+        key = c[:80].lower()
+        if key not in seen:
+            seen.add(key)
+            deduped.append(c)
+
+    return deduped[:_MAX_CLAUSES]
 
 
 def retrieve_for_clause(
@@ -117,7 +176,33 @@ def retrieve_for_clause(
     top_k: int = 5,
     domains: str | None = None,
 ) -> list["RetrievedChunk"]:
-    return retrieval_service.retrieve(query=clause, top_k=top_k, filters=domains)
+    """Retrieve regulation chunks relevant to a clause.
+
+    If the best retrieval score is below 0.55, rewrite the clause into a more
+    technical query and retry once to improve coverage.
+    """
+    chunks = retrieval_service.retrieve(query=clause, top_k=top_k, filters=domains)
+    if not chunks:
+        return chunks
+
+    best_score = max((getattr(c, "score", 0) for c in chunks), default=0)
+    if best_score < 0.55:
+        # Rewrite clause as technical keyword query and retry
+        keywords = " ".join(
+            w for w in re.split(r"\W+", clause) if len(w) > 3
+        )[:200]
+        retry_chunks = retrieval_service.retrieve(query=keywords, top_k=top_k, filters=domains)
+        if retry_chunks:
+            # Merge: keep unique chunks, prefer higher-score version
+            seen_ids: set[str] = {getattr(c, "chunk_id", str(i)) for i, c in enumerate(chunks)}
+            for rc in retry_chunks:
+                rid = getattr(rc, "chunk_id", "")
+                if rid not in seen_ids:
+                    chunks.append(rc)
+                    seen_ids.add(rid)
+            chunks.sort(key=lambda c: getattr(c, "score", 0), reverse=True)
+            chunks = chunks[:top_k]
+    return chunks
 
 
 def process_single_clause(
@@ -130,14 +215,75 @@ def process_single_clause(
 ) -> dict:
     """Process one clause: retrieve relevant regulations then check compliance.
 
-    Returns a dict with keys: index, chunks, finding (may be None on LLM failure).
+    Returns a dict with keys:
+      - index: clause position (for ordering)
+      - chunks: list of RetrievedChunk (for source events)
+      - finding: dict with title/desc/status/clause_ref/confidence (may be None on LLM failure)
+
     Designed to run inside asyncio.to_thread() for parallel execution.
+    The finding now includes a 'source_refs' list linking back to the chunks
+    that informed the verdict, enabling the frontend to correlate sources with findings.
     """
     chunks = retrieve_for_clause(clause, retrieval_service, top_k, domains)
     finding = check_clause_compliance(clause, chunks, client)
+    if finding is not None:
+        # Attach source references so the frontend can link finding ↔ sources
+        finding["source_refs"] = [
+            {
+                "standard": getattr(c, "doc_title", "") or getattr(c, "doc_name", ""),
+                "clause": getattr(c, "section_title", "") or "",
+                "score": round(float(getattr(c, "score", 0)), 3),
+            }
+            for c in chunks[:3]
+        ]
     return {"index": index, "chunks": chunks, "finding": finding}
 
 
+async def run_clauses_streaming(
+    clauses: list[str],
+    retrieval_service: "KnowledgeRetrievalService",
+    client: "BaseLLMClient",
+    top_k: int = 5,
+    domains: str | None = None,
+):
+    """Process all clauses concurrently and yield each result as it completes.
+
+    Unlike the old gather()-based approach, this uses asyncio.Queue so that
+    findings are emitted to the SSE stream immediately when each clause
+    finishes — the user sees results progressively rather than waiting for
+    the slowest clause before seeing any output.
+
+    Yields dicts with keys: index, chunks, finding (same schema as
+    process_single_clause, plus a sentinel {"_done": True} at the end).
+    """
+    queue: asyncio.Queue[dict] = asyncio.Queue()
+    total = len(clauses)
+
+    async def _worker(clause: str, i: int) -> None:
+        """Run one clause in a thread and push the result into the queue."""
+        try:
+            result = await asyncio.to_thread(
+                process_single_clause,
+                clause, i, retrieval_service, client, top_k, domains,
+            )
+        except Exception as exc:
+            logger.warning("Clause {} processing failed: {}", i, exc)
+            result = {"index": i, "chunks": [], "finding": None}
+        await queue.put(result)
+
+    # Launch all workers concurrently
+    tasks = [asyncio.create_task(_worker(clause, i)) for i, clause in enumerate(clauses)]
+
+    received = 0
+    while received < total:
+        result = await queue.get()
+        yield result
+        received += 1
+
+    # Wait for all tasks to complete (they should already be done by now)
+    await asyncio.gather(*tasks, return_exceptions=True)
+
+
 async def run_clauses_parallel(
     clauses: list[str],
     retrieval_service: "KnowledgeRetrievalService",
@@ -145,31 +291,15 @@ async def run_clauses_parallel(
     top_k: int = 5,
     domains: str | None = None,
 ) -> list[dict]:
-    """Run all clauses through retrieve+gap-check in parallel.
+    """Legacy batch API kept for backward compatibility.
 
-    Results are returned in the original clause order even though processing
-    is concurrent. Exceptions in individual clauses are caught and returned as
-    dicts with finding=None so the stream continues for remaining clauses.
-
-    Both retrieval_service and client must be thread-safe — they are shared
-    across all asyncio.to_thread() calls without locking.
+    Collects all streaming results and returns them sorted by clause index.
+    New code should use run_clauses_streaming() directly.
     """
-    tasks = [
-        asyncio.to_thread(
-            process_single_clause,
-            clause, i, retrieval_service, client, top_k, domains,
-        )
-        for i, clause in enumerate(clauses)
-    ]
-    raw = await asyncio.gather(*tasks, return_exceptions=True)
-    results = []
-    for i, r in enumerate(raw):
-        if isinstance(r, Exception):
-            logger.warning("Clause {} processing failed: {}", i, r)
-            results.append({"index": i, "chunks": [], "finding": None})
-        else:
-            results.append(r)
-    return results
+    results: list[dict] = []
+    async for result in run_clauses_streaming(clauses, retrieval_service, client, top_k, domains):
+        results.append(result)
+    return sorted(results, key=lambda r: r["index"])
 
 
 def check_clause_compliance(
@@ -177,6 +307,15 @@ def check_clause_compliance(
     chunks: list["RetrievedChunk"],
     client: "BaseLLMClient",
 ) -> dict | None:
+    """Check whether a business clause complies with the retrieved regulations.
+
+    The prompt explicitly instructs the LLM to:
+    - extract clause_ref from the retrieved text (not invent it)
+    - include a confidence score (0-1) reflecting how well the retrieved
+      chunks cover the clause topic
+
+    Returns None only when the LLM call fails after all retries.
+    """
     reg_context = "\n".join(
         f"[{i+1}] {c.doc_title} {c.section_title or ''}: {c.text[:300]}"
         for i, c in enumerate(chunks[:5])
@@ -186,14 +325,17 @@ def check_clause_compliance(
         "complies with the retrieved regulations.\n\n"
         f"Business clause:\n{clause}\n\n"
         f"Retrieved regulations:\n{reg_context}\n\n"
-        "Return JSON:\n"
+        "Return JSON with these exact fields:\n"
         "{\n"
         '  "status": "ok" | "warn" | "risk",\n'
         '  "title": "Short finding title (max 30 chars)",\n'
         '  "desc": "Description (50-120 chars)",\n'
-        '  "clause_ref": "Regulation clause reference e.g. Art.9.1 or Sec.3.1"\n'
+        '  "clause_ref": "Exact clause/article reference copied from the retrieved text above, '
+        'e.g. Art.9.1 or Sec.3.1. Use null if no specific clause number appears in the retrieved text.",\n'
+        '  "confidence": 0.0-1.0  // how well the retrieved context covers this clause topic\n'
         "}\n"
         "status: ok=compliant, warn=gap exists, risk=critical/missing\n"
+        "IMPORTANT: copy clause_ref verbatim from the retrieved text; do NOT invent references.\n"
         "Return ONLY the JSON object."
     )
 
@@ -216,7 +358,10 @@ def check_clause_compliance(
                 "title": str(result.get("title", "Compliance finding")),
                 "desc": str(result.get("desc", "")),
                 "status": result.get("status", "info"),
-                "clause_ref": result.get("clause_ref"),
+                # None if LLM correctly found no clause number in retrieved text
+                "clause_ref": result.get("clause_ref") or None,
+                # Confidence score helps frontend show retrieval quality indicator
+                "confidence": float(result.get("confidence", 0.5)),
             }
     except (ValueError, TypeError) as exc:
         logger.warning("Gap check JSON parse failed: {}", exc)
@@ -368,3 +513,58 @@ def generate_suggestions(
     except (ValueError, TypeError) as exc:
         logger.warning("generate_suggestions JSON parse failed: {}", exc)
     return fallback
+
+
+def detect_cross_clause_conflicts(
+    findings: list[dict],
+    client: "BaseLLMClient",
+) -> list[dict]:
+    """Detect contradictions and missing cross-references across all findings.
+
+    Runs a single LLM call after all per-clause findings are collected.
+    Returns a list of conflict dicts: {type, finding_a, finding_b, desc}.
+    Returns an empty list on LLM failure so the caller can proceed without it.
+    """
+    if len(findings) < 2:
+        # Need at least 2 findings to compare
+        return []
+
+    findings_text = "\n".join(
+        f"[{i+1}] [{f['status'].upper()}] {f['title']}: {f['desc']}"
+        + (f" (Ref: {f['clause_ref']})" if f.get("clause_ref") else "")
+        for i, f in enumerate(findings)
+    )
+    prompt = (
+        "You are a compliance expert. Review the following compliance findings from the same document "
+        "and identify any cross-clause issues:\n\n"
+        f"Findings:\n{findings_text}\n\n"
+        "Return JSON array of conflicts (empty array [] if none found):\n"
+        "[\n"
+        "  {\n"
+        '    "type": "contradiction" | "missing_ref" | "cumulative_risk",\n'
+        '    "finding_a": <1-based index>,\n'
+        '    "finding_b": <1-based index or null>,\n'
+        '    "desc": "Brief description of the cross-clause issue (max 100 chars)"\n'
+        "  }\n"
+        "]\n"
+        "Return ONLY the JSON array."
+    )
+    try:
+        response = client.chat([{"role": "user", "content": prompt}], max_tokens=600)
+        if not response.is_success:
+            return []
+        result = _extract_json(response.content)
+        if isinstance(result, list):
+            return [
+                {
+                    "type": str(c.get("type", "contradiction")),
+                    "finding_a": int(c.get("finding_a", 0)),
+                    "finding_b": c.get("finding_b"),
+                    "desc": str(c.get("desc", "")),
+                }
+                for c in result
+                if isinstance(c, dict)
+            ]
+    except Exception as exc:
+        logger.warning("detect_cross_clause_conflicts failed: {}", exc)
+    return []

backend/app/application/documents/services.py

@@ -526,10 +526,28 @@ class DocumentCommandService:
                     logger.warning("临时文件清理失败: {}", temp_path)
 
     def delete(self, doc_id: str) -> bool:
-        """Delete document record, binary file, and vector chunks."""
+        """Delete document record, binary file, and vector chunks.
+
+        Handles two cases:
+        - Normal docs: have a metadata record in the document repository.
+        - Milvus-only (synthetic) docs: visible in management-list because they
+          have Milvus vectors but no JSON/PG metadata record. We still clean up
+          the Milvus chunks so the document disappears from the list.
+        """
         document = self.document_repository.get(doc_id)
         if not document:
+            # No metadata record — might be a Milvus-only synthetic document.
+            # Attempt vector cleanup directly; treat as success if any chunks deleted.
+            try:
+                deleted_count = self.vector_index.delete_by_document(doc_id)
+                if deleted_count > 0:
+                    logger.info("Deleted Milvus-only doc (no metadata record): doc_id={} chunks={}", doc_id, deleted_count)
+                    return True
+            except Exception as exc:
+                logger.warning("Milvus-only delete failed for doc_id={}: {}", doc_id, exc)
             return False
+
+        # Normal doc: clean up binary, vectors, artifacts, processing records, metadata.
         try:
             self.binary_store.delete(document.object_name)
         except Exception:
@@ -627,13 +645,16 @@ class DocumentQueryService:
             result.append(doc)
 
         # Surface Milvus-only docs that have no metadata record at all.
+        # MinIO almost certainly has their binaries (they were uploaded), so
+        # set object_name to the sentinel "{doc_id}/" so the route marks
+        # has_file=True; the download endpoint will list MinIO to find the file.
         for doc_id, row in milvus_by_id.items():
             if doc_id not in meta_by_id:
                 synthetic = Document(
                     doc_id=doc_id,
                     doc_name=row.get("doc_title", doc_id),
                     file_name=row.get("doc_title", doc_id),
-                    object_name="",
+                    object_name=f"{doc_id}/",  # sentinel: MinIO prefix exists
                     content_type="",
                     size_bytes=0,
                     status=DocumentStatus.INDEXED,
@@ -646,9 +667,63 @@ class DocumentQueryService:
         result.sort(key=lambda d: d.updated_at, reverse=True)
         return result[:limit] if limit is not None else result
 
-    def download(self, doc_id: str) -> tuple[Document, bytes]:
-        """Handle download for the Document Query Service instance."""
+    def download(self, doc_id: str) -> tuple["Document", bytes]:
+        """Return the document record and its binary content from MinIO.
+
+        Fallback strategy for Milvus-only docs (no JSON/PG metadata record):
+          1. Try metadata repository first (normal path).
+          2. If metadata is missing, list MinIO objects with prefix ``{doc_id}/``
+             and synthesise a minimal Document from the first object found.
+             This handles documents whose metadata records were lost but whose
+             binary files are still in object storage.
+          3. If neither source has the file, raise FileNotFoundError.
+        """
+        from app.domain.documents import Document, DocumentStatus
+
         document = self.document_repository.get(doc_id)
-        if not document:
-            raise FileNotFoundError(f"文档不存在: {doc_id}")
+
+        if document and document.object_name and not document.object_name.endswith("/"):
+            # Normal doc with a concrete object_name — read directly.
+            return document, self.binary_store.read(document.object_name)
+
+        if document and not document.object_name:
+            raise FileNotFoundError(f"该文档无原始文件（仅含索引数据，无法下载）: {doc_id}")
+
+        if not document or document.object_name.endswith("/"):
+            # Metadata missing — try to find the file in MinIO by doc_id prefix.
+            try:
+                objects = self.binary_store.list_objects(prefix=f"{doc_id}/")
+                # Filter out artifact JSON files; prefer the source document.
+                candidates = [o for o in objects if not o.endswith(".json")]
+                if not candidates:
+                    candidates = objects  # fall back to all objects if only JSON found
+                if not candidates:
+                    raise FileNotFoundError(f"文档不存在（MinIO 和元数据均无记录）: {doc_id}")
+                object_name = candidates[0]
+                file_name = object_name.split("/", 1)[-1] if "/" in object_name else object_name
+                # Guess content type from extension.
+                ext = file_name.rsplit(".", 1)[-1].lower() if "." in file_name else ""
+                _ct_map = {
+                    "pdf": "application/pdf",
+                    "docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
+                    "doc": "application/msword",
+                    "txt": "text/plain",
+                }
+                content_type = _ct_map.get(ext, "application/octet-stream")
+                # Synthesise a minimal Document so the route can build the response.
+                document = Document(
+                    doc_id=doc_id,
+                    doc_name=file_name,
+                    file_name=file_name,
+                    object_name=object_name,
+                    content_type=content_type,
+                    size_bytes=0,
+                    status=DocumentStatus.INDEXED,
+                )
+                logger.info("MinIO fallback download: doc_id={} object={}", doc_id, object_name)
+            except FileNotFoundError:
+                raise
+            except Exception as exc:
+                raise FileNotFoundError(f"文档不存在: {doc_id}") from exc
+
         return document, self.binary_store.read(document.object_name)

backend/app/config/settings.py

@@ -133,6 +133,42 @@ class Settings(BaseSettings):
     reranker_api_key: str = Field(default="", description="Reranker API 密钥")
     reranker_top_k: int = Field(default=5, description="精排后保留的最终结果数量")
 
+    # ── HyDE (Hypothetical Document Embeddings) ──────────────────────────────
+    # When enabled, the agentic and standard RAG pipelines generate a short
+    # hypothetical answer before retrieval, then embed that text instead of the
+    # raw query. This closes the vocabulary gap between terse queries and longer
+    # document passages, typically improving recall by 15-30% on vague queries.
+    hyde_enabled: bool = Field(default=True, description="启用 HyDE 查询增强（假设文档嵌入）")
+    hyde_max_tokens: int = Field(default=200, description="HyDE 假设段落最大 token 数")
+    # Use a lightweight model for HyDE to reduce latency and cost.
+    # HyDE only needs a short plausible passage — a fast cheap model is sufficient.
+    # Leave empty to fall back to the main llm_provider / llm_model.
+    hyde_llm_provider: str = Field(default="", description="HyDE 专用 LLM 提供商（空则复用主 LLM）")
+    hyde_llm_model: str = Field(default="", description="HyDE 专用 LLM 模型（空则复用主 LLM）")
+
+    # ── Agentic RAG (P0-1) ───────────────────────────────────────────────────
+    # Controls the multi-step reasoning pipeline exposed at /agent/agentic/stream.
+    agentic_max_sub_queries: int = Field(
+        default=4,
+        description="Agentic 模式最大子查询分解数量（compare / multi_hop 意图触发）",
+    )
+    agentic_grounding_threshold: float = Field(
+        default=0.65,
+        description=(
+            "引文锚定 fast-path 阈值：avg_score > 此值且 chunks ≥ 3 时跳过 LLM grounding check，"
+            "直接判定为充分；降低此值可让更多问题触发 LLM 二次验证。"
+        ),
+    )
+    agentic_intent_max_tokens: int = Field(
+        default=200, description="意图分析步骤 LLM 最大 token 数"
+    )
+    agentic_plan_max_tokens: int = Field(
+        default=400, description="查询分解步骤 LLM 最大 token 数"
+    )
+    agentic_grounding_max_tokens: int = Field(
+        default=250, description="引文锚定步骤 LLM 最大 token 数"
+    )
+
     # Keep configuration setup explicit so runtime behavior is easy to reason about.
     milvus_index_type: str = Field(default="IVF_FLAT", description="Milvus索引类型")
     milvus_nlist: int = Field(default=128, description="Milvus nlist参数")

backend/app/infrastructure/llm/openai_compatible_answer_generator.py

@@ -9,10 +9,12 @@ from app.config.settings import settings
 from app.domain.conversation import AnswerGenerator, AnswerResult, AnswerSource
 from app.domain.retrieval import RetrievedChunk
 from app.services.llm.llm_factory import get_llm_client
+from app.services.rag.prompt_templates import PromptTemplates
 # Keep adapter behavior explicit so integration details remain easy to audit.
 
 
-PROMPT_TEMPLATES = {
+# Fallback system prompts used when no rich template matches.
+_FALLBACK_PROMPTS = {
     "default": "你是法规知识问答助手。请仅依据提供的上下文回答；如果上下文不足，明确说明。",
     "compliance_qa": "你是法规合规问答助手。优先引用给定法规原文，回答要准确、克制，并注明依据来源。",
 }
@@ -38,33 +40,80 @@ class OpenAICompatibleAnswerGenerator(AnswerGenerator):
         retrieved_chunks: list[RetrievedChunk],
         history: list[dict[str, str]] | None,
         prompt_template: str | None,
+        context_text: str | None = None,
+        context_filename: str | None = None,
     ) -> tuple[list[dict[str, str]], int]:
-        """Handle build messages for this module for the Open A I Compatible Answer Generator instance."""
-        system_prompt = PROMPT_TEMPLATES.get(prompt_template or "compliance_qa", PROMPT_TEMPLATES["default"])
+        """Build the message list to send to the LLM.
+
+        When context_text is provided the user's document is injected as a
+        dedicated section BEFORE the retrieved regulation chunks so the LLM
+        can reason about the document directly while still referencing regulations.
+        The retrieval step uses only the user's question, not the document text,
+        so embedding quality is preserved.
+
+        System prompt selection priority:
+          1. Rich template from PromptTemplates (compliance_qa / comparison /
+             compliance_check / clause_interpretation / …)
+          2. Fallback hardcoded prompt when no rich template matches.
+        """
+        # Look up the rich template first; fall back to simple hardcoded prompts.
+        tpl_name = prompt_template or "compliance_qa"
+        rich_tpl = PromptTemplates.get_template(tpl_name)
+        if rich_tpl:
+            system_prompt = rich_tpl.system_prompt
+        else:
+            system_prompt = _FALLBACK_PROMPTS.get(tpl_name, _FALLBACK_PROMPTS["default"])
         context_blocks = []
         context_tokens = 0
+
+        # ── User document context (if attached) ───────────────────────────────
+        if context_text and context_text.strip():
+            doc_label = f"附件文档：{context_filename}" if context_filename else "附件文档"
+            doc_block = f"[{doc_label}]\n{context_text.strip()}"
+            doc_tokens = self._estimate_tokens(doc_block)
+            # Reserve at most half the context budget for the user document
+            half_budget = settings.rag_max_context_tokens // 2
+            if doc_tokens > half_budget:
+                # Truncate document to fit half the budget
+                ratio = half_budget / doc_tokens
+                doc_block = doc_block[: int(len(doc_block) * ratio)] + "\n…（文档已截断）"
+                doc_tokens = half_budget
+            context_blocks.append(doc_block)
+            context_tokens += doc_tokens
+
+        # ── Retrieved regulation chunks ────────────────────────────────────────
+        remaining_budget = settings.rag_max_context_tokens - context_tokens
         for idx, chunk in enumerate(retrieved_chunks, start=1):
             block = (
-                f"[{idx}] 文档: {chunk.doc_title}\n"
+                f"[法规{idx}] 文档: {chunk.doc_title}\n"
                 f"章节: {chunk.section_title or '未标注'}\n"
                 f"页码: {chunk.page_start}" + (f"-{chunk.page_end}" if chunk.page_end and chunk.page_end != chunk.page_start else "") + "\n"
                 f"内容: {chunk.text}"
             )
             block_tokens = self._estimate_tokens(block)
-            if context_tokens + block_tokens > settings.rag_max_context_tokens:
+            if block_tokens > remaining_budget:
                 break
+            remaining_budget -= block_tokens
             context_tokens += block_tokens
             context_blocks.append(block)
+
         context = "\n\n".join(context_blocks)
         messages = [{"role": "system", "content": system_prompt}]
         for item in history or []:
             messages.append({"role": item["role"], "content": item["content"]})
-        messages.append(
-            {
-                "role": "user",
-                "content": f"问题：{query}\n\n参考上下文：\n{context}\n\n请在回答后给出简要引用编号。",
-            }
-        )
+
+        # Craft the user turn differently when a document is attached
+        if context_text and context_text.strip():
+            user_content = (
+                f"问题：{query}\n\n"
+                f"请先基于上方附件文档内容进行分析，再结合法规参考上下文给出合规评估。"
+                f"\n\n参考上下文：\n{context}\n\n"
+                f"请在回答中注明引用来源编号（如适用）。"
+            )
+        else:
+            user_content = f"问题：{query}\n\n参考上下文：\n{context}\n\n请在回答后给出简要引用编号。"
+
+        messages.append({"role": "user", "content": user_content})
         return messages, context_tokens
 
     def _is_context_truncated(self, *, retrieved_chunks: list[RetrievedChunk], context_tokens: int) -> bool:
@@ -112,6 +161,8 @@ class OpenAICompatibleAnswerGenerator(AnswerGenerator):
         provider: str | None = None,
         model: str | None = None,
         prompt_template: str | None = None,
+        context_text: str | None = None,
+        context_filename: str | None = None,
     ) -> AnswerResult:
         """Handle generate for the Open A I Compatible Answer Generator instance."""
         start = time.time()
@@ -120,6 +171,8 @@ class OpenAICompatibleAnswerGenerator(AnswerGenerator):
             retrieved_chunks=retrieved_chunks,
             history=history,
             prompt_template=prompt_template,
+            context_text=context_text,
+            context_filename=context_filename,
         )
         client = get_llm_client(provider=provider or settings.llm_provider, model=model or settings.llm_model)
         response = client.chat(messages)
@@ -147,6 +200,8 @@ class OpenAICompatibleAnswerGenerator(AnswerGenerator):
         provider: str | None = None,
         model: str | None = None,
         prompt_template: str | None = None,
+        context_text: str | None = None,
+        context_filename: str | None = None,
     ) -> Generator[dict, None, AnswerResult]:
         """Stream generate for the Open A I Compatible Answer Generator instance."""
         start = time.time()
@@ -155,6 +210,8 @@ class OpenAICompatibleAnswerGenerator(AnswerGenerator):
             retrieved_chunks=retrieved_chunks,
             history=history,
             prompt_template=prompt_template,
+            context_text=context_text,
+            context_filename=context_filename,
         )
         sources = [source.__dict__ for source in self._sources(retrieved_chunks)]
         yield {"event": "sources", "data": sources}

backend/app/infrastructure/storage/minio_binary_store.py

@@ -41,6 +41,10 @@ class MinioDocumentBinaryStore(DocumentBinaryStore):
             raise FileNotFoundError(f"对象不存在: {object_name}")
         return data
 
+    def list_objects(self, prefix: str = "") -> list[str]:
+        """List object names in the bucket that start with the given prefix."""
+        return self.client.list_objects(prefix=prefix)
+
     def delete(self, object_name: str) -> None:
         """Handle delete for the Minio Document Binary Store instance."""
         if not self.client.delete_object(object_name):

backend/app/infrastructure/vectorstore/cross_encoder_reranker.py

@@ -71,22 +71,48 @@ class OpenAICompatibleReranker(Reranker):
         return result
 
     def _call_reranker(self, query: str, texts: list[str]) -> list[float]:
-        """Call the reranker API and return a score per text."""
+        """Call the reranker API and return a score per text.
+
+        Tries TEI format first (POST /rerank with model+texts), then falls back
+        to Cohere/OpenAI format (POST /v1/rerank with model+documents).
+        Both formats now include the model name, which most gateways require.
+        """
         headers = {"Content-Type": "application/json"}
         if self._api_key:
             headers["Authorization"] = f"Bearer {self._api_key}"
 
-        # Try TEI format first: POST /rerank
-        payload = {"query": query, "texts": texts, "raw_scores": False, "return_text": False}
+        # TEI format: POST /rerank — include model name (required by gateway proxies)
+        payload = {
+            "model": self._model,
+            "query": query,
+            "texts": texts,
+            "raw_scores": False,
+            "return_text": False,
+        }
         url = f"{self._base_url}/rerank"
         resp = requests.post(url, json=payload, headers=headers, timeout=self._timeout)
 
-        if resp.status_code == 404:
-            # Fall back to Cohere / OpenAI-style: POST /v1/rerank
+        if resp.status_code in (404, 400):
+            # Gateway returned an error — try Cohere/OpenAI-style format as fallback.
+            logger.debug(
+                "TEI rerank returned {} — trying Cohere format. Body: {}",
+                resp.status_code,
+                resp.text[:200],
+            )
             payload_v1 = {"model": self._model, "query": query, "documents": texts}
             url = f"{self._base_url}/v1/rerank"
             resp = requests.post(url, json=payload_v1, headers=headers, timeout=self._timeout)
 
+        if not resp.ok:
+            # Surface a clear error message so callers can log it meaningfully.
+            try:
+                err_body = resp.json()
+                err_msg = err_body.get("error", {}).get("message", resp.text[:200])
+            except Exception:
+                err_msg = resp.text[:200]
+            resp.raise_for_status()  # raises HTTPError with status code
+            raise ValueError(err_msg)  # unreachable but satisfies type checker
+
         resp.raise_for_status()
         data = resp.json()
 

backend/app/schemas/rag.py

@@ -12,6 +12,11 @@ class RagChatRequest(BaseModel):
     top_k: int = 5
     session_id: Optional[str] = None
     filters: Optional[str] = None
+    # Optional document text to inject directly as LLM conversation context.
+    # When provided the document content is prepended to the query so the LLM
+    # can answer questions about it without requiring vector-store indexing.
+    context_text: Optional[str] = None
+    context_filename: Optional[str] = None
 
 
 class RetrievedDoc(BaseModel):

backend/app/services/llm/base_client.py

@@ -1,9 +1,16 @@
-"""Provide service-layer logic for base client."""
+"""Provide service-layer logic for base client.
+
+P0-0: ``LLMResponse`` now carries an optional ``tool_calls`` list so that any
+downstream code (agents, pipelines) can inspect and dispatch tool invocations
+without touching the provider-specific adapter layer.
+"""
 
 from abc import ABC, abstractmethod
 from dataclasses import dataclass, field
 from typing import List, Dict, Optional, Any
 from enum import Enum
+
+from app.services.llm.tool_types import Tool, ToolCall  # noqa: F401 – re-exported for callers
 # Keep provider-specific behavior explicit so debugging stays straightforward.
 
 
@@ -24,6 +31,8 @@ class LLMResponse:
     finish_reason: str = "stop"
     latency_ms: int = 0
     error: Optional[str] = None
+    # P0-0: populated when the model returns tool-call(s) instead of plain text.
+    tool_calls: List[ToolCall] = field(default_factory=list)
 
     @property
     def is_success(self) -> bool:
@@ -63,9 +72,19 @@ class BaseLLMClient(ABC):
         messages: List[Dict[str, str]],
         max_tokens: Optional[int] = None,
         temperature: Optional[float] = None,
+        tools: Optional[List["Tool"]] = None,
         **kwargs
     ) -> LLMResponse:
-        """Handle chat for the Base L L M Client instance."""
+        """Handle chat for the Base L L M Client instance.
+
+        Args:
+            messages: OpenAI-format message list.
+            max_tokens: Override config max_tokens when set.
+            temperature: Override config temperature when set.
+            tools: Optional list of Tool definitions to offer the model.
+                   When provided, the model may respond with tool_calls in the
+                   returned LLMResponse instead of (or in addition to) content.
+        """
         pass
 
     def complete(

backend/app/services/llm/deepseek_client.py

@@ -1,4 +1,8 @@
-"""Provide service-layer logic for deepseek client."""
+"""Provide service-layer logic for deepseek client.
+
+P0-0: ``chat()`` now accepts an optional ``tools`` list and parses ``tool_calls``
+from the model response so that callers can dispatch tool invocations.
+"""
 
 import time
 from typing import List, Dict, Optional
@@ -6,6 +10,7 @@ from loguru import logger
 import httpx
 
 from .base_client import BaseLLMClient, LLMResponse, LLMConfig, LLMProvider
+from .tool_types import Tool, ToolCall
 # Keep provider-specific behavior explicit so debugging stays straightforward.
 
 
@@ -46,13 +51,20 @@ class DeepSeekClient(BaseLLMClient):
         messages: List[Dict[str, str]],
         max_tokens: Optional[int] = None,
         temperature: Optional[float] = None,
+        tools: Optional[List[Tool]] = None,
         **kwargs
     ) -> LLMResponse:
-        """Handle chat for the Deep Seek Client instance."""
+        """Handle chat for the Deep Seek Client instance.
+
+        When ``tools`` is provided the request includes the tool definitions and
+        ``tool_choice="auto"``; any tool_calls returned by the model are parsed
+        into ``LLMResponse.tool_calls``.
+        """
+        import json
         start_time = time.time()
 
         try:
-            payload = {
+            payload: Dict = {
                 "model": self.config.model,
                 "messages": messages,
                 "max_tokens": max_tokens or self.config.max_tokens,
@@ -61,6 +73,11 @@ class DeepSeekClient(BaseLLMClient):
                 "stream": False
             }
 
+            # P0-0: inject tool definitions when provided.
+            if tools:
+                payload["tools"] = [t.to_openai_format() for t in tools]
+                payload["tool_choice"] = "auto"
+
             response = self._client.post("/chat/completions", json=payload)
             response.raise_for_status()
 
@@ -71,12 +88,24 @@ class DeepSeekClient(BaseLLMClient):
             choices = data.get("choices", [{}])
             message = choices[0].get("message", {})
 
+            # P0-0: parse tool_calls returned by the model.
+            raw_tool_calls = message.get("tool_calls") or []
+            parsed_tool_calls: List[ToolCall] = []
+            for tc in raw_tool_calls:
+                fn = tc.get("function", {})
+                try:
+                    args = json.loads(fn.get("arguments", "{}"))
+                except json.JSONDecodeError:
+                    args = {}
+                parsed_tool_calls.append(ToolCall(id=tc.get("id", ""), name=fn.get("name", ""), arguments=args))
+
             return LLMResponse(
-                content=message.get("content", ""),
+                content=message.get("content", "") or "",
                 model=data.get("model", self.config.model),
                 usage=data.get("usage", {}),
                 finish_reason=choices[0].get("finish_reason", "stop"),
-                latency_ms=latency_ms
+                latency_ms=latency_ms,
+                tool_calls=parsed_tool_calls,
             )
 
         except httpx.HTTPStatusError as e:

backend/app/services/llm/qwen_client.py

@@ -1,4 +1,8 @@
-"""Provide service-layer logic for qwen client."""
+"""Provide service-layer logic for qwen client.
+
+P0-0: ``chat()`` now accepts an optional ``tools`` list and parses ``tool_calls``
+from the model response so that callers can dispatch tool invocations.
+"""
 
 import time
 import json
@@ -7,6 +11,7 @@ from loguru import logger
 import httpx
 
 from .base_client import BaseLLMClient, LLMResponse, LLMConfig, LLMProvider
+from .tool_types import Tool, ToolCall
 # Keep provider-specific behavior explicit so debugging stays straightforward.
 
 
@@ -54,14 +59,20 @@ class QwenClient(BaseLLMClient):
         messages: List[Dict[str, str]],
         max_tokens: Optional[int] = None,
         temperature: Optional[float] = None,
+        tools: Optional[List[Tool]] = None,
         **kwargs
     ) -> LLMResponse:
-        """Handle chat for the Qwen Client instance."""
+        """Handle chat for the Qwen Client instance.
+
+        When ``tools`` is provided the request includes the tool definitions and
+        ``tool_choice="auto"``; any tool_calls returned by the model are parsed
+        into ``LLMResponse.tool_calls``.
+        """
         start_time = time.time()
 
         try:
             # Keep provider-specific behavior explicit so debugging stays straightforward.
-            payload = {
+            payload: Dict = {
                 "model": self.config.model,
                 "messages": messages,
                 "max_tokens": max_tokens or self.config.max_tokens,
@@ -70,6 +81,11 @@ class QwenClient(BaseLLMClient):
                 "stream": False
             }
 
+            # P0-0: inject tool definitions when provided.
+            if tools:
+                payload["tools"] = [t.to_openai_format() for t in tools]
+                payload["tool_choice"] = "auto"
+
             # Keep provider-specific behavior explicit so debugging stays straightforward.
             response = self._client.post("/chat/completions", json=payload)
             response.raise_for_status()
@@ -82,12 +98,24 @@ class QwenClient(BaseLLMClient):
             choices = data.get("choices", [{}])
             message = choices[0].get("message", {})
 
+            # P0-0: parse tool_calls returned by the model.
+            raw_tool_calls = message.get("tool_calls") or []
+            parsed_tool_calls: List[ToolCall] = []
+            for tc in raw_tool_calls:
+                fn = tc.get("function", {})
+                try:
+                    args = json.loads(fn.get("arguments", "{}"))
+                except json.JSONDecodeError:
+                    args = {}
+                parsed_tool_calls.append(ToolCall(id=tc.get("id", ""), name=fn.get("name", ""), arguments=args))
+
             return LLMResponse(
-                content=message.get("content", ""),
+                content=message.get("content", "") or "",
                 model=data.get("model", self.config.model),
                 usage=data.get("usage", {}),
                 finish_reason=choices[0].get("finish_reason", "stop"),
-                latency_ms=latency_ms
+                latency_ms=latency_ms,
+                tool_calls=parsed_tool_calls,
             )
 
         except httpx.HTTPStatusError as e:

backend/app/services/llm/tool_types.py

@@ -0,0 +1,80 @@
+"""Shared tool and tool-call type definitions for LLM function calling (P0-0).
+
+These types implement the OpenAI-compatible tool/function-calling interface so that
+any provider whose gateway supports the spec (DeepSeek, Qwen, etc.) can expose
+tools to the LLM and receive structured tool invocations in return.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class ToolCall:
+    """Represent a single tool invocation returned by the LLM.
+
+    The model fills in ``id``, ``name``, and ``arguments`` when it decides to call
+    a tool instead of (or in addition to) producing a text response.
+    """
+
+    # Unique identifier assigned by the model for this call.
+    id: str
+    # Name of the tool to invoke, matching the name registered in Tool.
+    name: str
+    # Parsed JSON arguments ready for direct use by the tool handler.
+    arguments: dict[str, Any] = field(default_factory=dict)
+
+
+@dataclass
+class ToolParameter:
+    """JSON-Schema–compatible parameter block for a tool definition."""
+
+    # Top-level schema type — always "object" for OpenAI-compatible tools.
+    type: str = "object"
+    # Map of parameter name → JSON-Schema property descriptor.
+    properties: dict[str, Any] = field(default_factory=dict)
+    # List of required parameter names.
+    required: list[str] = field(default_factory=list)
+
+
+@dataclass
+class Tool:
+    """Describe a callable tool that can be offered to the LLM.
+
+    Example usage::
+
+        search_tool = Tool(
+            name="search_regulations",
+            description="Search the compliance knowledge base for relevant regulation clauses.",
+            parameters=ToolParameter(
+                properties={"query": {"type": "string", "description": "Search query"}},
+                required=["query"],
+            ),
+        )
+        response = client.chat(messages, tools=[search_tool])
+    """
+
+    name: str
+    description: str
+    parameters: ToolParameter = field(default_factory=ToolParameter)
+
+    def to_openai_format(self) -> dict[str, Any]:
+        """Serialise this tool to the OpenAI-compatible function-calling schema.
+
+        The returned dict can be placed directly in the ``tools`` list of a chat
+        completions request without any further transformation.
+        """
+        return {
+            "type": "function",
+            "function": {
+                "name": self.name,
+                "description": self.description,
+                "parameters": {
+                    "type": self.parameters.type,
+                    "properties": self.parameters.properties,
+                    "required": self.parameters.required,
+                },
+            },
+        }

backend/app/shared/bootstrap.py

@@ -6,6 +6,7 @@ from functools import lru_cache
 from typing import Callable
 
 from app.application.agent import AgentConversationService, AgentSessionService
+from app.application.agent.agentic_service import AgenticConversationService
 from app.application.documents import DocumentCommandService, DocumentQueryService
 from app.application.knowledge import KnowledgeRetrievalService
 from app.application.perception.services import PerceptionService
@@ -365,6 +366,20 @@ def get_agent_session_service() -> AgentSessionService:
     return AgentSessionService(conversation_store=get_conversation_store())
 
 
+@lru_cache
+def get_agentic_conversation_service() -> AgenticConversationService:
+    """Return the Agentic RAG service (P0-1).
+
+    Uses the same retrieval, generation, and session infrastructure as the
+    standard chat service so no additional dependencies are required.
+    """
+    return AgenticConversationService(
+        retrieval_service=get_retrieval_service(),
+        answer_generator=OpenAICompatibleAnswerGenerator(),
+        conversation_store=get_conversation_store(),
+    )
+
+
 @lru_cache
 def get_celery_app():
     """Return the shared Celery application instance.

frontend/src/api/index.ts

@@ -73,6 +73,22 @@ export interface SSEMessage {
   text?: string;
   docs?: RetrievedDoc[];
   session_id?: string;
+  // ── P0-1 Agentic-mode thinking-step fields ────────────────────────────────
+  // Populated when type === 'thinking'; maps to the backend IntentResult /
+  // GroundingResult / retrieval step payloads emitted by AgenticConversationService.
+  step?: string;           // intent_analysis | query_planning | retrieving | grounding_check
+  status?: string;         // running | done
+  intent_type?: string;    // simple_qa | compare | multi_hop | ambiguous
+  requires_decomposition?: boolean;
+  reason?: string;
+  sub_queries?: string[];
+  query?: string;          // sub-query being retrieved
+  index?: number;          // 1-based sub-query index
+  total?: number;          // total sub-query count
+  found?: number;          // chunks found for this sub-query
+  retry?: boolean;         // true when this is a grounding-failure re-query
+  sufficient?: boolean;    // grounding check result
+  confidence?: number;     // grounding confidence 0–1
 }
 
 export async function streamSSE<TMessage extends SSEMessage>(

frontend/src/api/rag.ts

@@ -76,6 +76,27 @@ function parseSSEChunk(raw: string, onMessage: (data: SSEMessage) => void) {
       onMessage({ type: 'error', text: joined });
     } else if (eventName === 'status') {
       onMessage({ type: 'status', text: joined });
+    } else if (eventName === 'thinking') {
+      // P0-1: Agentic reasoning step events from /agent/agentic/stream
+      try {
+        const payload = JSON.parse(joined) as Record<string, unknown>;
+        onMessage({
+          type: 'thinking',
+          step: payload.step as string | undefined,
+          status: payload.status as string | undefined,
+          intent_type: payload.intent_type as string | undefined,
+          requires_decomposition: payload.requires_decomposition as boolean | undefined,
+          reason: payload.reason as string | undefined,
+          sub_queries: payload.sub_queries as string[] | undefined,
+          query: payload.query as string | undefined,
+          index: payload.index as number | undefined,
+          total: payload.total as number | undefined,
+          found: payload.found as number | undefined,
+          retry: payload.retry as boolean | undefined,
+          sufficient: payload.sufficient as boolean | undefined,
+          confidence: payload.confidence as number | undefined,
+        });
+      } catch { /* ignore */ }
     } else if (eventName === 'message') {
       // /rag/chat format: event:message + JSON body with type field
       try {
@@ -147,3 +168,73 @@ export async function ragChat(
 }
 
 export type { QuickQuestionsResponse, SSEMessage };
+
+/**
+ * P0-1 Agentic RAG chat — calls /agent/agentic/stream which runs the full
+ * intent-analysis → query-planning → retrieval → grounding-check → answer pipeline.
+ *
+ * The onMessage callback receives the same event types as ragChat plus
+ * ``type: 'thinking'`` events that carry live reasoning-step progress.
+ */
+export async function agenticChat(
+  query: string,
+  topK: number = 5,
+  onMessage: (data: SSEMessage) => void,
+  onError?: (error: Error) => void,
+  onComplete?: () => void,
+  filters?: string,
+  sessionId?: string,
+  signal?: AbortSignal,
+  contextText?: string,
+  contextFilename?: string,
+): Promise<void> {
+  try {
+    const response = await fetch(`${AGENT_API_BASE}/agent/agentic/stream`, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+        Accept: 'text/event-stream',
+        ...(getToken() ? { Authorization: `Bearer ${getToken()}` } : {}),
+      },
+      body: JSON.stringify({
+        query,
+        top_k: topK,
+        ...(filters ? { filters } : {}),
+        ...(sessionId ? { session_id: sessionId } : {}),
+        ...(contextText ? { context_text: contextText, context_filename: contextFilename ?? '' } : {}),
+      }),
+      signal,
+    });
+
+    if (!response.ok || !response.body) {
+      throw new Error(`HTTP error! status: ${response.status}`);
+    }
+
+    const reader = response.body.getReader();
+    const decoder = new TextDecoder();
+    let buffer = '';
+
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+
+      buffer += decoder.decode(value, { stream: true });
+      const parts = buffer.split('\n\n');
+      buffer = parts.pop() || '';
+      parseSSEChunk(parts.join('\n\n'), onMessage);
+    }
+
+    if (buffer.trim()) {
+      parseSSEChunk(buffer, onMessage);
+    }
+
+    if (onComplete) {
+      onComplete();
+    }
+  } catch (error) {
+    if (error instanceof DOMException && error.name === 'AbortError') return;
+    if (onError) {
+      onError(error instanceof Error ? error : new Error(String(error)));
+    }
+  }
+}

frontend/src/contexts/PageStateContext.tsx

@@ -59,6 +59,8 @@ export interface ComplianceSourceEvent {
   score: number;
   status: string;
   full_content: string;
+  /** Index of the clause this source was retrieved for (for source↔finding linking) */
+  clause_index?: number;
 }
 
 export interface ComplianceFindingEvent {
@@ -66,6 +68,17 @@ export interface ComplianceFindingEvent {
   desc: string;
   status: 'ok' | 'warn' | 'risk';
   clause_ref?: string;
+  /** LLM confidence that retrieved context covers the clause topic (0–1) */
+  confidence?: number;
+  /** Top-3 regulation chunks that informed this finding */
+  source_refs?: Array<{ standard: string; clause: string; score: number }>;
+}
+
+export interface ComplianceConflict {
+  type: 'contradiction' | 'missing_ref' | 'cumulative_risk';
+  finding_a: number;
+  finding_b: number | null;
+  desc: string;
 }
 
 export interface ComplianceActionItem {
@@ -103,6 +116,10 @@ export interface ComplianceState {
   analysisId: string | null;
   isReadOnly: boolean;
   activeFindingId: string | null;
+  /** Real-time per-clause progress {done, total} */
+  progress: { done: number; total: number } | null;
+  /** Cross-clause conflicts detected after all findings complete */
+  conflicts: ComplianceConflict[];
 }
 
 const COMPLIANCE_INIT: ComplianceState = {
@@ -117,6 +134,8 @@ const COMPLIANCE_INIT: ComplianceState = {
   analysisId: null,
   isReadOnly: false,
   activeFindingId: null,
+  progress: null,
+  conflicts: [],
 };
 
 // ── Perception types ──────────────────────────────────────────────────────────

frontend/src/contexts/index.ts

@@ -12,6 +12,7 @@ export type {
   ComplianceStatus,
   ComplianceSourceEvent,
   ComplianceFindingEvent,
+  ComplianceConflict,
   ComplianceDonePayload,
   ComplianceMeta,
   ComplianceActionItem,

frontend/src/locales/en.ts

@@ -238,6 +238,36 @@ export interface Translations {
     citationsHeader: string;
     citationsEmpty: string;
     apiError: string;
+    // ── Agentic mode ─────────────────────────────────────────────────────────
+    agenticMode: string;
+    agenticModeHint: string;
+    agentThinking: string;
+    agentDone: string;
+    stepSuffix: string;
+    stepIntentAnalysis: string;
+    stepQueryPlanning: string;
+    stepRetrieving: string;
+    stepGrounding: string;
+    intentSimpleQa: string;
+    intentCompare: string;
+    intentMultiHop: string;
+    intentAmbiguous: string;
+    intentNeedsDecomposition: string;
+    subQueriesCountSuffix: string;
+    chunksFoundSuffix: string;
+    retryLabel: string;
+    groundingSufficient: string;
+    groundingInsufficient: string;
+    // ── Document attachment in interface ─────────────────────────────────────
+    attachBtn: string;
+    attachExtracting: string;
+    attachReady: string;
+    attachError: string;
+    attachClearLabel: string;
+    attachContextBadge: string;
+    attachAccept: string;
+    attachTruncated: string;
+    attachErrorMsg: string;
   };
 }
 
@@ -480,5 +510,35 @@ export const en: Translations = {
     citationsHeader: 'Sources',
     citationsEmpty: 'Citations will appear here after a response is generated.',
     apiError: 'Could not reach the RAG API. Please check the backend.',
+    // ── Agentic mode ─────────────────────────────────────────────────────────
+    agenticMode: 'Agentic mode',
+    agenticModeHint: 'Intent · Planning · Retrieval · Grounding',
+    agentThinking: 'Agent reasoning…',
+    agentDone: 'Reasoning complete',
+    stepSuffix: 'steps',
+    stepIntentAnalysis: 'Intent analysis',
+    stepQueryPlanning: 'Query planning',
+    stepRetrieving: 'Knowledge retrieval',
+    stepGrounding: 'Citation grounding',
+    intentSimpleQa: 'Simple Q&A',
+    intentCompare: 'Comparison',
+    intentMultiHop: 'Multi-hop',
+    intentAmbiguous: 'Ambiguous',
+    intentNeedsDecomposition: 'Decomposed',
+    subQueriesCountSuffix: 'sub-queries',
+    chunksFoundSuffix: 'chunks',
+    retryLabel: '(retry) ',
+    groundingSufficient: '✓ Sufficient',
+    groundingInsufficient: '⚠ Re-queried',
+    // ── Document context attachment ───────────────────────────────────────────
+    attachBtn: 'Attach document as context',
+    attachExtracting: 'Extracting text…',
+    attachReady: 'Context loaded',
+    attachError: 'Extraction failed',
+    attachClearLabel: 'Clear',
+    attachContextBadge: 'Doc context',
+    attachAccept: '.pdf,.docx,.doc,.txt,.md',
+    attachTruncated: '(truncated to 8 000 chars)',
+    attachErrorMsg: 'Could not extract text from this file.',
   },
 };

frontend/src/locales/zh.ts

@@ -239,5 +239,35 @@ export const zh: Translations = {
     citationsHeader: '引用来源',
     citationsEmpty: '生成回答后，引用来源将显示在此处。',
     apiError: '无法连接到 RAG API，请检查后端服务。',
+    // ── Agentic mode ─────────────────────────────────────────────────────────
+    agenticMode: 'Agentic 模式',
+    agenticModeHint: '意图分析 · 查询分解 · 迭代检索 · 引文锚定',
+    agentThinking: 'Agent 推理中…',
+    agentDone: '推理完成',
+    stepSuffix: '步',
+    stepIntentAnalysis: '意图分析',
+    stepQueryPlanning: '查询分解',
+    stepRetrieving: '知识检索',
+    stepGrounding: '引文锚定',
+    intentSimpleQa: '单跳问答',
+    intentCompare: '对比分析',
+    intentMultiHop: '多跳推理',
+    intentAmbiguous: '模糊查询',
+    intentNeedsDecomposition: '需分解',
+    subQueriesCountSuffix: '个子查询',
+    chunksFoundSuffix: '条',
+    retryLabel: '(补充) ',
+    groundingSufficient: '✓ 充分',
+    groundingInsufficient: '⚠ 补充检索',
+    // ── Document context attachment ───────────────────────────────────────────
+    attachBtn: '上传文档作为对话上下文',
+    attachExtracting: '正在提取文本…',
+    attachReady: '上下文已加载',
+    attachError: '提取失败',
+    attachClearLabel: '清除',
+    attachContextBadge: '文档上下文',
+    attachAccept: '.pdf,.docx,.doc,.txt,.md',
+    attachTruncated: '（已截断至 8000 字符）',
+    attachErrorMsg: '无法从该文件提取文本，请检查文件格式。',
   },
 };

frontend/src/pages/Compliance/CompliancePage.tsx

@@ -1,6 +1,6 @@
 import { useState, useRef, useEffect } from 'react';
 import { useLanguage } from '../../contexts/LanguageContext';
-import { Search, Plus, AlertTriangle, Download, MessageSquare, ChevronDown } from 'lucide-react';
+import { Search, Plus, Download, MessageSquare, ChevronDown, AlertTriangle } from 'lucide-react';
 import { Topbar } from '../../components/layout/Topbar';
 import { NewAnalysisModal } from './NewAnalysisModal';
 import { useComplianceAnalysis } from './useComplianceAnalysis';
@@ -39,81 +39,8 @@ function formatTs(iso: string) {
   } catch { return iso; }
 }
 
-// ── Chat state for a single finding ─────────────────────────────────────────
-interface ChatMsg { id: number; role: 'user' | 'assistant'; content: string }
-
-function useFindingChat() {
-  const [open, setOpen] = useState(false);
-  const [findingIdx, setFindingIdx] = useState<number | null>(null);
-  const [messages, setMessages] = useState<ChatMsg[]>([]);
-  const [input, setInput] = useState('');
-  const [loading, setLoading] = useState(false);
-  const abortRef = useRef<AbortController | null>(null);
-
-  function openFor(idx: number, finding: FindingEvent) {
-    setFindingIdx(idx);
-    setOpen(true);
-    setMessages([{
-      id: 0,
-      role: 'assistant',
-      content: `I'm reviewing finding: **${finding.title}**\n\n${finding.desc}${finding.clause_ref ? `\n\nRef: ${finding.clause_ref}` : ''}\n\nHow can I help?`,
-    }]);
-    setInput('');
-  }
-
-  function close() { setOpen(false); abortRef.current?.abort(); }
-
-  async function send(segmentContext: string) {
-    if (!input.trim() || loading) return;
-    const q = input.trim();
-    setInput('');
-    const userMsg: ChatMsg = { id: Date.now(), role: 'user', content: q };
-    const assistantId = Date.now() + 1;
-    setMessages(m => [...m, userMsg, { id: assistantId, role: 'assistant', content: '' }]);
-    setLoading(true);
-
-    const ctrl = new AbortController();
-    abortRef.current = ctrl;
-
-    try {
-      const res = await fetch(`/api/v1/compliance/chat/${findingIdx ?? 0}`, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json', ...authHeader() },
-        body: JSON.stringify({ query: q, segment_context: segmentContext }),
-        signal: ctrl.signal,
-      });
-      if (!res.body) { setLoading(false); return; }
-      const reader = res.body.getReader();
-      const dec = new TextDecoder();
-      let buf = '';
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        buf += dec.decode(value, { stream: true });
-        const blocks = buf.split('\n\n');
-        buf = blocks.pop() ?? '';
-        for (const block of blocks) {
-          const dl = block.split('\n').find(l => l.startsWith('data: '));
-          if (!dl) continue;
-          try {
-            const j = JSON.parse(dl.slice(6));
-            if (j.type === 'chunk' && j.text) {
-              setMessages(m => m.map(msg => msg.id === assistantId ? { ...msg, content: msg.content + j.text } : msg));
-            }
-          } catch { /* skip */ }
-        }
-      }
-    } catch (e: unknown) {
-      if (e instanceof Error && e.name === 'AbortError') return;
-    } finally {
-      setLoading(false);
-    }
-  }
-
-  return { open, findingIdx, messages, input, setInput, loading, openFor, close, send };
-}
-
-function _FindingChatDrawerWrapper({
+/** Wrapper that resolves findingIndex → findingId from the saved analysis, then renders FindingChatDrawer. */
+function FindingChatDrawerWrapper({
   analysisId,
   findingIndex,
   finding,
@@ -128,7 +55,7 @@ function _FindingChatDrawerWrapper({
 
   useEffect(() => {
     fetch(`/api/v1/compliance/history/${analysisId}`, {
-      headers: { Authorization: `Bearer ${localStorage.getItem('auth_token') ?? ''}` },
+      headers: authHeader(),
     })
       .then(r => r.json())
       .then((data: { findings?: Array<{ seq: number; id: string }> }) => {
@@ -153,8 +80,8 @@ export function CompliancePage() {
   const [showModal, setShowModal] = useState(false);
   const [showExportMenu, setShowExportMenu] = useState(false);
   const { state, run, reset } = useComplianceAnalysis();
-  const chat = useFindingChat();
-  const [drawerFindingIdx, setDrawerFindingIdx] = useState<number | null>(null);
+  // drawerFinding holds {index, finding} for the currently-open FindingChatDrawer
+  const [drawerFinding, setDrawerFinding] = useState<{ idx: number; finding: FindingEvent } | null>(null);
 
   const { setComplianceState } = usePageState();
   const { t } = useLanguage();
@@ -198,6 +125,8 @@ export function CompliancePage() {
       analysisId: data.id,
       isReadOnly: true,
       activeFindingId: null,
+      progress: null,
+      conflicts: [],
     });
   }
 
@@ -258,12 +187,6 @@ export function CompliancePage() {
     setShowExportMenu(false);
   }
 
-  // ── Chat context (finding desc + clause_ref as segment context) ──────────
-  const activeFinding = chat.findingIdx !== null ? state.findings[chat.findingIdx] : null;
-  const chatContext = activeFinding
-    ? `Finding: ${activeFinding.title}\n${activeFinding.desc}${activeFinding.clause_ref ? `\nRef: ${activeFinding.clause_ref}` : ''}`
-    : '';
-
   return (
     <div className="compliance-page" style={{ position: 'relative' }}>
       <Topbar
@@ -457,6 +380,25 @@ export function CompliancePage() {
             <div className="comp-col findings-col">
               <div className="col-header">
                 Findings {state.findings.length > 0 && `(${state.findings.length})`}
+                {/* Real per-clause progress bar during streaming */}
+                {isStreaming && state.progress && state.progress.total > 0 && (
+                  <span style={{
+                    marginLeft: 8, fontSize: 10, color: 'var(--muted)',
+                    display: 'inline-flex', alignItems: 'center', gap: 6,
+                  }}>
+                    <span style={{
+                      display: 'inline-block', width: 60, height: 4,
+                      background: 'var(--border)', borderRadius: 2, overflow: 'hidden',
+                    }}>
+                      <span style={{
+                        display: 'block', height: '100%',
+                        width: `${Math.round((state.progress.done / state.progress.total) * 100)}%`,
+                        background: 'var(--accent)', transition: 'width 0.3s ease',
+                      }} />
+                    </span>
+                    {state.progress.done}/{state.progress.total}
+                  </span>
+                )}
               </div>
 
               {state.findings.length === 0 && isStreaming && (
@@ -472,30 +414,85 @@ export function CompliancePage() {
                     <span className={`status ${f.status}`}>{STATUS_LABEL[f.status] ?? f.status}</span>
                   </div>
                   <p className="finding-desc">{f.desc}</p>
+
+                  {/* Source refs: which retrieved chunks informed this finding */}
+                  {f.source_refs && f.source_refs.length > 0 && (
+                    <div style={{ marginTop: 4, display: 'flex', flexWrap: 'wrap', gap: 4 }}>
+                      {f.source_refs.map((sr, si) => (
+                        <span key={si} style={{
+                          fontSize: 10, padding: '1px 6px',
+                          background: 'var(--bg)', border: '1px solid var(--border)',
+                          borderRadius: 4, color: 'var(--muted)',
+                        }} title={sr.clause}>
+                          📄 {sr.standard ? sr.standard.slice(0, 20) : '—'}
+                          {sr.score > 0 && ` · ${Math.round(sr.score * 100)}%`}
+                        </span>
+                      ))}
+                    </div>
+                  )}
+
                   <div style={{ display: 'flex', alignItems: 'center', justifyContent: 'space-between', marginTop: 6 }}>
-                    {f.clause_ref && (
-                      <div style={{ fontSize: 11, color: 'var(--muted)' }}>Ref: {f.clause_ref}</div>
-                    )}
-                    <button
-                      className="btn sm"
-                      style={{ marginLeft: 'auto', fontSize: 11, padding: '3px 8px', gap: 4 }}
-                      onClick={() => chat.openFor(i, f)}
-                    >
-                      <MessageSquare size={11} />{t.compliance.askAIBtn}
-                    </button>
-                    {state.analysisId && (
+                    <div style={{ display: 'flex', alignItems: 'center', gap: 8 }}>
+                      {f.clause_ref && (
+                        <div style={{ fontSize: 11, color: 'var(--muted)' }}>Ref: {f.clause_ref}</div>
+                      )}
+                      {/* Confidence dot: green ≥0.7, amber 0.4–0.7, red <0.4 */}
+                      {f.confidence !== undefined && (
+                        <span
+                          style={{
+                            fontSize: 10, color: 'var(--muted)',
+                            display: 'inline-flex', alignItems: 'center', gap: 3,
+                          }}
+                          title={`Retrieval confidence: ${Math.round(f.confidence * 100)}%`}
+                        >
+                          <span style={{
+                            width: 6, height: 6, borderRadius: '50%',
+                            background: f.confidence >= 0.7 ? '#22c55e' : f.confidence >= 0.4 ? '#f59e0b' : '#ef4444',
+                          }} />
+                          {Math.round(f.confidence * 100)}%
+                        </span>
+                      )}
+                    </div>
+                    {/* Single consolidated chat button — only when analysis is saved */}
+                    {state.analysisId ? (
                       <button
                         className="btn sm"
-                        onClick={() => setDrawerFindingIdx(i)}
-                        style={{ marginTop: 6 }}
+                        style={{ marginLeft: 'auto', fontSize: 11, padding: '3px 8px', gap: 4 }}
+                        onClick={() => setDrawerFinding({ idx: i, finding: f })}
                       >
-                        💬 {t.compliance.chatBtn}
+                        <MessageSquare size={11} />{t.compliance.chatBtn}
                       </button>
+                    ) : (
+                      /* Fallback for unsaved analyses: show disabled chat hint */
+                      <span style={{ marginLeft: 'auto', fontSize: 10, color: 'var(--muted)' }}>
+                        {t.compliance.askAIBtn}
+                      </span>
                     )}
                   </div>
                 </div>
               ))}
 
+              {/* Cross-clause conflicts panel */}
+              {state.conflicts && state.conflicts.length > 0 && (
+                <div className="card" style={{ borderLeft: '3px solid #f59e0b', marginTop: 8 }}>
+                  <div className="card-header" style={{ display: 'flex', alignItems: 'center', gap: 6 }}>
+                    <AlertTriangle size={12} color="#f59e0b" />
+                    <span style={{ fontSize: 12, fontWeight: 600 }}>Cross-Clause Issues ({state.conflicts.length})</span>
+                  </div>
+                  {state.conflicts.map((c, ci) => (
+                    <div key={ci} style={{ fontSize: 11, color: 'var(--muted)', padding: '4px 0', borderTop: ci ? '1px solid var(--border)' : 'none' }}>
+                      <span style={{
+                        fontWeight: 600,
+                        color: c.type === 'contradiction' ? '#ef4444' : c.type === 'cumulative_risk' ? '#f59e0b' : 'var(--fg)',
+                      }}>
+                        [{c.type.replace('_', ' ')}]
+                      </span>
+                      {' '}Finding #{c.finding_a}{c.finding_b ? ` ↔ #${c.finding_b}` : ''}: {c.desc}
+                    </div>
+                  ))}
+                </div>
+              )}
+
               {/* Conclusion */}
               {isDone && state.done && (
                 <div className="card conclusion-box">
@@ -540,92 +537,18 @@ export function CompliancePage() {
             </div>
           </div>
 
-          {/* ── Finding Chat Side Panel ────────────────────────────────── */}
-          {chat.open && (
-            <div style={{
-              position: 'fixed', right: 0, top: 0, bottom: 0, width: 400,
-              background: 'var(--surface)', borderLeft: '1px solid var(--border)',
-              display: 'flex', flexDirection: 'column', zIndex: 200,
-              boxShadow: '-8px 0 32px rgba(0,0,0,.12)',
-            }}>
-              {/* Header */}
-              <div style={{ padding: '16px 20px', borderBottom: '1px solid var(--border)', display: 'flex', alignItems: 'center', justifyContent: 'space-between' }}>
-                <div>
-                  <div style={{ fontSize: 13, fontWeight: 600 }}>{t.compliance.chatSidebarHeader}</div>
-                  <div style={{ fontSize: 11, color: 'var(--muted)', marginTop: 2 }}>
-                    Finding #{(chat.findingIdx ?? 0) + 1} · {activeFinding?.title}
-                  </div>
-                </div>
-                <button
-                  onClick={chat.close}
-                  style={{ background: 'none', border: 'none', cursor: 'pointer', color: 'var(--muted)', padding: 4 }}
-                >✕</button>
-              </div>
-
-              {/* Messages */}
-              <div style={{ flex: 1, overflowY: 'auto', padding: '16px 20px', display: 'flex', flexDirection: 'column', gap: 12 }}>
-                {chat.messages.map(msg => (
-                  <div key={msg.id} style={{ display: 'flex', gap: 10, flexDirection: msg.role === 'user' ? 'row-reverse' : 'row' }}>
-                    {msg.role === 'assistant' && (
-                      <div style={{ width: 28, height: 28, borderRadius: 8, background: 'var(--accent)', display: 'flex', alignItems: 'center', justifyContent: 'center', flexShrink: 0, fontSize: 11, color: '#fff', fontWeight: 700 }}>AI</div>
-                    )}
-                    <div style={{
-                      maxWidth: '82%', padding: '10px 14px', borderRadius: 10, fontSize: 13, lineHeight: 1.6, whiteSpace: 'pre-wrap',
-                      background: msg.role === 'user' ? 'var(--accent)' : 'var(--bg)',
-                      color: msg.role === 'user' ? '#fff' : 'var(--fg)',
-                      border: msg.role === 'assistant' ? '1px solid var(--border)' : 'none',
-                    }}>{msg.content}</div>
-                  </div>
-                ))}
-                {chat.loading && (
-                  <div style={{ display: 'flex', gap: 10 }}>
-                    <div style={{ width: 28, height: 28, borderRadius: 8, background: 'var(--accent)', display: 'flex', alignItems: 'center', justifyContent: 'center', flexShrink: 0, fontSize: 11, color: '#fff', fontWeight: 700 }}>AI</div>
-                    <div style={{ padding: '10px 14px', borderRadius: 10, border: '1px solid var(--border)', background: 'var(--bg)', fontSize: 13, color: 'var(--muted)' }}>
-                      {t.compliance.chatThinking}
-                    </div>
-                  </div>
-                )}
-              </div>
-
-              {/* Quick questions */}
-              <div style={{ padding: '8px 20px', display: 'flex', flexWrap: 'wrap', gap: 6 }}>
-                {[t.compliance.quickQ1, t.compliance.quickQ2, t.compliance.quickQ3].map(q => (
-                  <button key={q} onClick={() => chat.setInput(q)}
-                    style={{ padding: '4px 10px', fontSize: 11, background: 'var(--bg)', border: '1px solid var(--border)', borderRadius: 6, cursor: 'pointer', color: 'var(--muted)' }}>
-                    {q}
-                  </button>
-                ))}
-              </div>
-
-              {/* Input */}
-              <div style={{ padding: '12px 20px', borderTop: '1px solid var(--border)', display: 'flex', gap: 8 }}>
-                <input
-                  value={chat.input}
-                  onChange={e => chat.setInput(e.target.value)}
-                  onKeyDown={e => { if (e.key === 'Enter' && !e.shiftKey) { e.preventDefault(); chat.send(chatContext); } }}
-                  placeholder={t.compliance.chatPlaceholder}
-                  style={{ flex: 1, padding: '9px 12px', fontSize: 13, background: 'var(--bg)', border: '1px solid var(--border)', borderRadius: 8, color: 'var(--fg)', outline: 'none' }}
-                />
-                <button
-                  className="btn primary"
-                  onClick={() => chat.send(chatContext)}
-                  disabled={!chat.input.trim() || chat.loading}
-                  style={{ padding: '9px 14px' }}
-                >{t.compliance.sendBtn}</button>
-              </div>
-            </div>
-          )}
-          {drawerFindingIdx !== null && state.analysisId && (
-            <_FindingChatDrawerWrapper
+          {/* ── Finding Chat Drawer (single consolidated UI) ───────────── */}
+          {drawerFinding !== null && state.analysisId && (
+            <FindingChatDrawerWrapper
               analysisId={state.analysisId}
-              findingIndex={drawerFindingIdx}
+              findingIndex={drawerFinding.idx}
               finding={{
-                title: state.findings[drawerFindingIdx]?.title ?? '',
-                desc: state.findings[drawerFindingIdx]?.desc ?? '',
-                status: state.findings[drawerFindingIdx]?.status ?? 'ok',
-                clause_ref: state.findings[drawerFindingIdx]?.clause_ref,
+                title: drawerFinding.finding.title,
+                desc: drawerFinding.finding.desc,
+                status: drawerFinding.finding.status,
+                clause_ref: drawerFinding.finding.clause_ref,
               }}
-              onClose={() => setDrawerFindingIdx(null)}
+              onClose={() => setDrawerFinding(null)}
             />
           )}
         </>

frontend/src/pages/Compliance/useComplianceAnalysis.ts

@@ -14,9 +14,10 @@ import type {
   ComplianceSourceEvent,
   ComplianceFindingEvent,
   ComplianceDonePayload,
+  ComplianceConflict,
 } from '../../contexts';
 
-export type { ComplianceMeta, ComplianceState, ComplianceSourceEvent as SourceEvent, ComplianceFindingEvent as FindingEvent, ComplianceDonePayload as DonePayload };
+export type { ComplianceMeta, ComplianceState, ComplianceSourceEvent as SourceEvent, ComplianceFindingEvent as FindingEvent, ComplianceDonePayload as DonePayload, ComplianceConflict };
 export type { ComplianceActionItem as ActionItem } from '../../contexts';
 export type AnalysisStatus = import('../../contexts').ComplianceStatus;
 export type AnalysisMeta = ComplianceMeta;
@@ -38,6 +39,8 @@ const INITIAL_STATE: ComplianceState = {
   errorText: '',
   analysisId: null,
   isReadOnly: false,
+  progress: null,
+  conflicts: [],
 };
 
 export function useComplianceAnalysis() {
@@ -92,6 +95,9 @@ export function useComplianceAnalysis() {
 
             if (j.type === 'stage') {
               setState(s => ({ ...s, stageLabel: j.label ?? '', stageKey: j.stage ?? '' }));
+            } else if (j.type === 'progress') {
+              // Real per-clause progress update from backend
+              setState(s => ({ ...s, progress: { done: j.done ?? 0, total: j.total ?? 0 } }));
             } else if (j.type === 'source') {
               const src: ComplianceSourceEvent = {
                 standard: j.standard ?? '',
@@ -99,6 +105,7 @@ export function useComplianceAnalysis() {
                 score: j.score ?? 0,
                 status: j.status ?? 'retrieved',
                 full_content: j.full_content ?? '',
+                clause_index: j.clause_index,
               };
               setState(s => ({ ...s, sources: [...s.sources, src] }));
             } else if (j.type === 'finding') {
@@ -107,8 +114,13 @@ export function useComplianceAnalysis() {
                 desc: j.desc ?? '',
                 status: j.status ?? 'info',
                 clause_ref: j.clause_ref,
+                confidence: j.confidence,
+                source_refs: j.source_refs,
               };
               setState(s => ({ ...s, findings: [...s.findings, finding] }));
+            } else if (j.type === 'conflicts') {
+              // Cross-clause conflicts detected after all findings finish
+              setState(s => ({ ...s, conflicts: j.items ?? [] }));
             } else if (j.type === 'done') {
               const payload: ComplianceDonePayload = {
                 conclusion: j.conclusion ?? '',

frontend/src/pages/Docs/DocsPage.tsx

@@ -20,6 +20,7 @@ interface Doc {
   sizeBytes: number;
   summary?: string;
   version?: string;
+  hasFile: boolean;
 }
 
 const STATUS_FILTERS = ['All', 'Ready', 'Processing', 'Failed', 'Pending'];
@@ -102,6 +103,7 @@ export function DocsPage() {
           sizeBytes: (item.size_bytes as number) ?? 0,
           summary: item.summary as string | undefined,
           version: item.version as string | undefined,
+          hasFile: item.has_file !== false,
         })));
         setLoading(false);
       })
@@ -130,11 +132,21 @@ export function DocsPage() {
   }
 
   // ── Download ─────────────────────────────────────────────────────────────
-  function downloadDoc(id: string, name: string) {
-    const a = document.createElement('a');
-    a.href = `/api/v1/documents/download/${id}`;
-    a.download = name;
-    a.click();
+  async function downloadDoc(id: string, name: string) {
+    try {
+      const resp = await fetch(`/api/v1/documents/download/${id}`, { headers: authHeader() });
+      if (!resp.ok) throw new Error(`下载失败: ${resp.status}`);
+      const blob = await resp.blob();
+      const url = URL.createObjectURL(blob);
+      const a = document.createElement('a');
+      a.href = url;
+      a.download = name;
+      a.click();
+      URL.revokeObjectURL(url);
+    } catch (err) {
+      console.error('Download failed', err);
+      alert(String(err));
+    }
   }
 
   // ── Retry (re-process failed doc) ────────────────────────────────────────
@@ -289,11 +301,13 @@ export function DocsPage() {
                   <span className="cell-mono">{formatSize(d.sizeBytes)}</span>
                   <span className="cell-muted">{d.type}</span>
                   <span className="row-actions">
-                    {/* Download */}
+                    {/* Download — disabled for Milvus-only docs that have no binary file */}
                     <button
                       className="text-link"
-                      title={t.docs.titleDownload}
+                      title={d.hasFile ? t.docs.titleDownload : '无原始文件'}
                       onClick={() => downloadDoc(d.id, d.name)}
+                      disabled={!d.hasFile}
+                      style={!d.hasFile ? { opacity: 0.3, cursor: 'not-allowed' } : undefined}
                     >
                       <Download size={12} />
                     </button>

frontend/src/pages/Docs/UploadModal.tsx

@@ -222,7 +222,7 @@ export function UploadModal({ onClose, onComplete }: Props) {
         <button className="modal-close" onClick={onClose} aria-label="Close" disabled={submitting}><X size={14} /></button>
 
         {/* ── Left panel: upload form ── */}
-        <div className="modal-panel">
+        <div className="modal-panel" style={{ overflowY: 'auto' }}>
           <div className="modal-eyebrow">Upload documents</div>
           <div className="modal-title">Stage files for parsing and indexing.</div>
           <p className="modal-lead">PDF, DOCX, TXT — one per API call, processed sequentially.</p>
@@ -254,7 +254,7 @@ export function UploadModal({ onClose, onComplete }: Props) {
           </div>
 
           {files.length > 0 && (
-            <div className="staged-files">
+            <div className="staged-files" style={{ maxHeight: 220, overflowY: 'auto', overflowX: 'hidden' }}>
               {files.map((f, i) => {
                 const isDone   = doneCount > i;
                 const isActive = submitting && currentFileIdx === i;

frontend/src/pages/RagChat/RagChatPage.tsx

@@ -1,9 +1,11 @@
 import { useRef, useEffect, useCallback, useState } from 'react';
 import { Topbar } from '../../components/layout/Topbar';
-import { Send, Download } from 'lucide-react';
+import { Send, Download, Zap, Paperclip, X, FileText, AlertCircle } from 'lucide-react';
 import { usePageState } from '../../contexts';
 import type { RagCitation } from '../../contexts';
 import { useLanguage } from '../../contexts/LanguageContext';
+import { agenticChat } from '../../api/rag';
+import type { SSEMessage } from '../../api/index';
 
 const TOKEN_KEY = 'auth_token';
 function authHeader(): Record<string, string> {
@@ -11,6 +13,46 @@ function authHeader(): Record<string, string> {
   return t ? { Authorization: `Bearer ${t}` } : {};
 }
 
+// ── Document context state ─────────────────────────────────────────────────────
+
+interface DocContext {
+  filename: string;
+  text: string;
+  charCount: number;
+  truncated: boolean;
+  /** 'extracting' while the backend is parsing; 'ready' when text is available; 'error' on failure */
+  status: 'extracting' | 'ready' | 'error';
+  errorMsg?: string;
+}
+
+// ── Agentic-mode types ────────────────────────────────────────────────────────
+
+interface ThinkingStep {
+  id: string;
+  step: string;
+  status: 'running' | 'done';
+  intent_type?: string;
+  reason?: string;
+  requires_decomposition?: boolean;
+  sub_queries?: string[];
+  query?: string;
+  index?: number;
+  total?: number;
+  found?: number;
+  sufficient?: boolean;
+  confidence?: number;
+  retry?: boolean;
+}
+
+const STEP_ICONS: Record<string, string> = {
+  intent_analysis: '🔍',
+  query_planning: '📋',
+  retrieving: '📚',
+  grounding_check: '🔗',
+};
+
+// ── Helpers ───────────────────────────────────────────────────────────────────
+
 // Map a raw source doc from the backend "retrieved" event to our Citation shape.
 function mapSource(s: Record<string, unknown>, idx: number): RagCitation {
   const rawScore = typeof s.score === 'number' ? s.score : 0;
@@ -69,10 +111,72 @@ export function RagChatPage() {
   const [streaming, setStreaming] = useState(ragStreamingRef.current);
   const [quickPrompts, setQuickPrompts] = useState<string[]>(MOCK_QUICK);
 
+  // P0-1 Agentic mode state
+  const [agenticMode, setAgenticMode] = useState(false);
+  const [thinkingSteps, setThinkingSteps] = useState<ThinkingStep[]>([]);
+  const [thinkingExpanded, setThinkingExpanded] = useState(true);
+
+  // ── Document context state ─────────────────────────────────────────────────
+  // Holds the extracted text from the attached file; sent to the backend as
+  // conversation context on every message while it is set.
+  const [docContext, setDocContext] = useState<DocContext | null>(null);
+  const fileInputRef = useRef<HTMLInputElement>(null);
+
   const bottomRef = useRef<HTMLDivElement>(null);
   const citRailRef = useRef<HTMLDivElement>(null);
   const citItemRefs = useRef<Record<number, HTMLDivElement | null>>({});
 
+  // ── Document context helpers ───────────────────────────────────────────────
+
+  /** Upload file to /rag/upload-context, extract its text, store as context. */
+  async function handleFileAttach(file: File) {
+    setDocContext({ filename: file.name, text: '', charCount: 0, truncated: false, status: 'extracting' });
+
+    const fd = new FormData();
+    fd.append('file', file);
+
+    try {
+      const res = await fetch('/api/v1/rag/upload-context', {
+        method: 'POST',
+        headers: authHeader(),
+        body: fd,
+      });
+      if (!res.ok) {
+        const errText = await res.text().catch(() => t.ragchat.attachErrorMsg);
+        setDocContext(prev => prev ? { ...prev, status: 'error', errorMsg: errText.slice(0, 120) } : null);
+        return;
+      }
+      const data = await res.json();
+      setDocContext({
+        filename: data.filename ?? file.name,
+        text: data.text ?? '',
+        charCount: data.char_count ?? 0,
+        truncated: data.truncated ?? false,
+        status: 'ready',
+      });
+    } catch (err) {
+      setDocContext(prev => prev
+        ? { ...prev, status: 'error', errorMsg: String(err).slice(0, 120) }
+        : null
+      );
+    }
+  }
+
+  function handleFileInputChange(e: React.ChangeEvent<HTMLInputElement>) {
+    const file = e.target.files?.[0];
+    if (file) void handleFileAttach(file);
+    // Reset so the same file can be re-selected
+    e.target.value = '';
+  }
+
+  function handleFileDrop(e: React.DragEvent<HTMLDivElement>) {
+    e.preventDefault();
+    const file = Array.from(e.dataTransfer.files).find(f =>
+      /\.(pdf|docx?|txt|md)$/i.test(f.name)
+    );
+    if (file) void handleFileAttach(file);
+  }
+
   // Fetch quick questions from backend on mount (only once per session)
   useEffect(() => {
     fetch('/api/v1/rag/quick-questions', { headers: authHeader() })
@@ -102,9 +206,17 @@ export function RagChatPage() {
 
   async function send(text?: string) {
     const q = (text ?? inputDraft).trim();
-    if (!q || ragStreamingRef.current) return;
+    // Block send while a document is still being extracted
+    if (!q || ragStreamingRef.current || docContext?.status === 'extracting') return;
+
     setRagState(s => ({ ...s, inputDraft: '' }));
 
+    // Show document context badge in user message bubble when active
+    const docPrefix = docContext?.status === 'ready'
+      ? `📄 ${docContext.filename}\n`
+      : '';
+    const displayQuery = docPrefix + q;
+
     const userMsgId = Date.now().toString();
     const assistantId = (Date.now() + 1).toString();
 
@@ -112,7 +224,7 @@ export function RagChatPage() {
       ...s,
       messages: [
         ...s.messages,
-        { id: userMsgId, role: 'user', text: q },
+        { id: userMsgId, role: 'user', text: displayQuery },
         { id: assistantId, role: 'assistant', text: '' },
       ],
       citations: [],
@@ -122,100 +234,211 @@ export function RagChatPage() {
     setStreaming(true);
     setHighlightedCit(null);
 
+    // P0-1: reset thinking panel for new query
+    if (agenticMode) {
+      setThinkingSteps([]);
+      setThinkingExpanded(true);
+    }
+
     const ctrl = new AbortController();
     ragAbortRef.current = ctrl;
 
-    try {
-      const body: Record<string, unknown> = { query: q, top_k: 5 };
-      if (sessionId) body.session_id = sessionId;
-
-      const res = await fetch('/api/v1/rag/chat', {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json', ...authHeader() },
-        body: JSON.stringify(body),
-        signal: ctrl.signal,
-      });
-
-      if (!res.body) throw new Error('No stream');
-      const reader = res.body.getReader();
-      const dec = new TextDecoder();
-      let buffer = '';
+    if (agenticMode) {
+      // ── Agentic path ────────────────────────────────────────────────────
       const newCitations: RagCitation[] = [];
 
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done) break;
-        buffer += dec.decode(value, { stream: true });
+      const handleMessage = (msg: SSEMessage) => {
+        if (msg.type === 'session') {
+          if (msg.session_id) setRagState(s => ({ ...s, sessionId: msg.session_id! }));
 
-        const blocks = buffer.split('\n\n');
-        buffer = blocks.pop() ?? '';
-
-        for (const block of blocks) {
-          const dataLine = block.split('\n').find(l => l.startsWith('data: '));
-          if (!dataLine) continue;
-          const raw = dataLine.slice(6).trim();
-          if (!raw) continue;
-          try {
-            const j = JSON.parse(raw);
-
-            if (j.type === 'session') {
-              if (j.session_id) setRagState(s => ({ ...s, sessionId: j.session_id }));
-
-            } else if (j.type === 'retrieved' && Array.isArray(j.docs)) {
-              const mapped = j.docs.map((d: Record<string, unknown>, i: number) => mapSource(d, i + 1));
-              newCitations.push(...mapped);
-              setRagState(s => ({ ...s, citations: [...mapped] }));
-
-            } else if (j.type === 'chunk' && j.text) {
-              setRagState(s => ({
-                ...s,
-                messages: s.messages.map(msg =>
-                  msg.id === assistantId
-                    ? { ...msg, text: msg.text + (j.text as string) }
-                    : msg
-                ),
-              }));
-
-            } else if (j.type === 'done') {
-              setRagState(s => ({
-                ...s,
-                messages: s.messages.map(msg => {
-                  if (msg.id !== assistantId) return msg;
-                  const refs = [...new Set(
-                    [...msg.text.matchAll(/\[(\d+)\]/g)].map(r => parseInt(r[1], 10))
-                  )].filter(n => n >= 1 && n <= newCitations.length);
-                  return { ...msg, citationRefs: refs };
-                }),
-              }));
-              break;
-
-            } else if (j.type === 'error') {
-              setRagState(s => ({
-                ...s,
-                messages: s.messages.map(msg =>
-                  msg.id === assistantId
-                    ? { ...msg, text: `Error: ${j.text ?? 'Unknown error'}` }
-                    : msg
-                ),
-              }));
+        } else if (msg.type === 'thinking') {
+          // Build a stable step id so we can upsert running→done transitions.
+          const stepId = `${msg.step}-${msg.retry ? 'retry' : (msg.index ?? 0)}`;
+          setThinkingSteps(prev => {
+            const idx = prev.findIndex(s => s.id === stepId);
+            const stepObj: ThinkingStep = {
+              id: stepId,
+              step: msg.step ?? '',
+              status: (msg.status as 'running' | 'done') ?? 'running',
+              intent_type: msg.intent_type,
+              reason: msg.reason,
+              sub_queries: msg.sub_queries,
+              query: msg.query,
+              index: msg.index,
+              total: msg.total,
+              found: msg.found,
+              sufficient: msg.sufficient,
+              confidence: msg.confidence,
+              retry: msg.retry,
+            };
+            if (idx >= 0) {
+              const updated = [...prev];
+              updated[idx] = stepObj;
+              return updated;
             }
-          } catch { /* malformed JSON chunk, skip */ }
+            return [...prev, stepObj];
+          });
+
+        } else if (msg.type === 'retrieved' && Array.isArray(msg.docs)) {
+          const mapped = (msg.docs as unknown as Record<string, unknown>[]).map((d, i) => mapSource(d, i + 1));
+          newCitations.push(...mapped);
+          setRagState(s => ({ ...s, citations: [...mapped] }));
+
+        } else if (msg.type === 'chunk' && msg.text) {
+          setRagState(s => ({
+            ...s,
+            messages: s.messages.map(m =>
+              m.id === assistantId ? { ...m, text: m.text + msg.text! } : m
+            ),
+          }));
+
+        } else if (msg.type === 'done') {
+          setThinkingExpanded(false);
+          setRagState(s => ({
+            ...s,
+            messages: s.messages.map(m => {
+              if (m.id !== assistantId) return m;
+              const refs = [...new Set(
+                [...m.text.matchAll(/\[(\d+)\]/g)].map(r => parseInt(r[1], 10))
+              )].filter(n => n >= 1 && n <= newCitations.length);
+              return { ...m, citationRefs: refs };
+            }),
+          }));
+
+        } else if (msg.type === 'error') {
+          setRagState(s => ({
+            ...s,
+            messages: s.messages.map(m =>
+              m.id === assistantId ? { ...m, text: `Error: ${msg.text ?? 'Unknown error'}` } : m
+            ),
+          }));
         }
+      };
+
+      try {
+        await agenticChat(
+          q, 5, handleMessage,
+          (err) => {
+            setRagState(s => ({
+              ...s,
+              messages: s.messages.map(m =>
+                m.id === assistantId ? { ...m, text: t.ragchat.apiError } : m
+              ),
+            }));
+            console.error('agenticChat error:', err);
+          },
+          undefined,
+          undefined,
+          sessionId ?? undefined,
+          ctrl.signal,
+          // Pass document context to agentic pipeline
+          docContext?.status === 'ready' ? docContext.text : undefined,
+          docContext?.status === 'ready' ? docContext.filename : undefined,
+        );
+      } finally {
+        ragStreamingRef.current = false;
+        setStreaming(false);
       }
-    } catch (e: unknown) {
-      if (e instanceof Error && e.name !== 'AbortError') {
-        setRagState(s => ({
-          ...s,
-          messages: s.messages.map(msg =>
-            msg.id === assistantId
-              ? { ...msg, text: t.ragchat.apiError }
-              : msg
-          ),
-        }));
+
+    } else {
+      // ── Standard RAG path (unchanged) ───────────────────────────────────
+      try {
+        const body: Record<string, unknown> = { query: q, top_k: 5 };
+        if (sessionId) body.session_id = sessionId;
+        // Inject document text as conversation context when a file is attached
+        if (docContext?.status === 'ready') {
+          body.context_text = docContext.text;
+          body.context_filename = docContext.filename;
+        }
+
+        const res = await fetch('/api/v1/rag/chat', {
+          method: 'POST',
+          headers: { 'Content-Type': 'application/json', ...authHeader() },
+          body: JSON.stringify(body),
+          signal: ctrl.signal,
+        });
+
+        if (!res.body) throw new Error('No stream');
+        const reader = res.body.getReader();
+        const dec = new TextDecoder();
+        let buffer = '';
+        const newCitations: RagCitation[] = [];
+
+        while (true) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          buffer += dec.decode(value, { stream: true });
+
+          const blocks = buffer.split('\n\n');
+          buffer = blocks.pop() ?? '';
+
+          for (const block of blocks) {
+            const dataLine = block.split('\n').find(l => l.startsWith('data: '));
+            if (!dataLine) continue;
+            const raw = dataLine.slice(6).trim();
+            if (!raw) continue;
+            try {
+              const j = JSON.parse(raw);
+
+              if (j.type === 'session') {
+                if (j.session_id) setRagState(s => ({ ...s, sessionId: j.session_id }));
+
+              } else if (j.type === 'retrieved' && Array.isArray(j.docs)) {
+                const mapped = j.docs.map((d: Record<string, unknown>, i: number) => mapSource(d, i + 1));
+                newCitations.push(...mapped);
+                setRagState(s => ({ ...s, citations: [...mapped] }));
+
+              } else if (j.type === 'chunk' && j.text) {
+                setRagState(s => ({
+                  ...s,
+                  messages: s.messages.map(msg =>
+                    msg.id === assistantId
+                      ? { ...msg, text: msg.text + (j.text as string) }
+                      : msg
+                  ),
+                }));
+
+              } else if (j.type === 'done') {
+                setRagState(s => ({
+                  ...s,
+                  messages: s.messages.map(msg => {
+                    if (msg.id !== assistantId) return msg;
+                    const refs = [...new Set(
+                      [...msg.text.matchAll(/\[(\d+)\]/g)].map(r => parseInt(r[1], 10))
+                    )].filter(n => n >= 1 && n <= newCitations.length);
+                    return { ...msg, citationRefs: refs };
+                  }),
+                }));
+                break;
+
+              } else if (j.type === 'error') {
+                setRagState(s => ({
+                  ...s,
+                  messages: s.messages.map(msg =>
+                    msg.id === assistantId
+                      ? { ...msg, text: `Error: ${j.text ?? 'Unknown error'}` }
+                      : msg
+                  ),
+                }));
+              }
+            } catch { /* malformed JSON chunk, skip */ }
+          }
+        }
+      } catch (e: unknown) {
+        if (e instanceof Error && e.name !== 'AbortError') {
+          setRagState(s => ({
+            ...s,
+            messages: s.messages.map(msg =>
+              msg.id === assistantId
+                ? { ...msg, text: t.ragchat.apiError }
+                : msg
+            ),
+          }));
+        }
+      } finally {
+        ragStreamingRef.current = false;
+        setStreaming(false);
       }
-    } finally {
-      ragStreamingRef.current = false;
-      setStreaming(false);
     }
   }
 
@@ -254,7 +477,126 @@ export function RagChatPage() {
 
         {/* ── Chat main ── */}
         <div className="chat-main">
-          <div className="messages">
+          {/* P0-1: Agentic Thinking Panel — shown when agentic mode is active */}
+          {agenticMode && thinkingSteps.length > 0 && (
+            <div style={{
+              margin: '0 0 4px 0',
+              border: '1px solid var(--border)',
+              borderRadius: 8,
+              background: streaming ? 'var(--surface)' : 'var(--surface-2, var(--surface))',
+              overflow: 'hidden',
+              transition: 'max-height 0.4s ease',
+              flexShrink: 0,
+            }}>
+              {/* Panel header — clickable to collapse/expand */}
+              <button
+                onClick={() => setThinkingExpanded(x => !x)}
+                style={{
+                  width: '100%',
+                  display: 'flex',
+                  alignItems: 'center',
+                  gap: 6,
+                  padding: '6px 12px',
+                  background: 'none',
+                  border: 'none',
+                  cursor: 'pointer',
+                  fontSize: 12,
+                  color: streaming ? 'var(--accent, #6366f1)' : 'var(--success-fg, #16a34a)',
+                  textAlign: 'left',
+                }}
+              >
+                <span>{streaming ? '⚙' : '✓'}</span>
+                <span style={{ fontWeight: 600 }}>
+                  {streaming
+                    ? t.ragchat.agentThinking
+                    : `${t.ragchat.agentDone} · ${thinkingSteps.filter(s => s.status === 'done').length} ${t.ragchat.stepSuffix}`
+                  }
+                </span>
+                <span style={{ marginLeft: 'auto', fontSize: 10 }}>{thinkingExpanded ? '▲' : '▼'}</span>
+              </button>
+              {/* Step list */}
+              {thinkingExpanded && (
+                <div style={{ padding: '0 12px 8px' }}>
+                  {thinkingSteps.map(step => {
+                    const stepLabels: Record<string, string> = {
+                      intent_analysis: t.ragchat.stepIntentAnalysis,
+                      query_planning:  t.ragchat.stepQueryPlanning,
+                      retrieving:      t.ragchat.stepRetrieving,
+                      grounding_check: t.ragchat.stepGrounding,
+                    };
+                    const intentLabels: Record<string, string> = {
+                      simple_qa:  t.ragchat.intentSimpleQa,
+                      compare:    t.ragchat.intentCompare,
+                      multi_hop:  t.ragchat.intentMultiHop,
+                      ambiguous:  t.ragchat.intentAmbiguous,
+                    };
+                    return (
+                      <div key={step.id} style={{
+                        display: 'flex',
+                        alignItems: 'flex-start',
+                        gap: 6,
+                        fontSize: 12,
+                        padding: '3px 0',
+                        color: step.status === 'done' ? 'var(--fg)' : 'var(--muted)',
+                      }}>
+                        <span style={{ width: 16, textAlign: 'center', flexShrink: 0 }}>
+                          {step.status === 'running'
+                            ? <span style={{ animation: 'spin 1s linear infinite', display: 'inline-block' }}>⟳</span>
+                            : (STEP_ICONS[step.step] ?? '·')
+                          }
+                        </span>
+                        <span>
+                          <strong>{stepLabels[step.step] ?? step.step}</strong>
+                          {/* Intent analysis detail */}
+                          {step.step === 'intent_analysis' && step.status === 'done' && step.intent_type && (
+                            <span style={{ marginLeft: 6, color: 'var(--muted)' }}>
+                              → {intentLabels[step.intent_type] ?? step.intent_type}
+                              {step.requires_decomposition && ` · ${t.ragchat.intentNeedsDecomposition}`}
+                            </span>
+                          )}
+                          {/* Query planning detail */}
+                          {step.step === 'query_planning' && step.status === 'done' && step.sub_queries && (
+                            <span style={{ marginLeft: 6, color: 'var(--muted)' }}>
+                              → {step.sub_queries.length} {t.ragchat.subQueriesCountSuffix}
+                            </span>
+                          )}
+                          {/* Retrieval detail */}
+                          {step.step === 'retrieving' && (
+                            <span style={{ marginLeft: 6, color: 'var(--muted)', wordBreak: 'break-all' }}>
+                              {step.total && step.total > 1 && `[${step.index}/${step.total}] `}
+                              {step.retry && t.ragchat.retryLabel}
+                              {step.query && step.query.length > 50
+                                ? step.query.slice(0, 50) + '…'
+                                : step.query}
+                              {step.status === 'done' && step.found !== undefined && (
+                                <span style={{ color: step.found > 0 ? 'var(--success-fg, #16a34a)' : 'var(--warning, #ca8a04)' }}>
+                                  {' '}· {step.found} {t.ragchat.chunksFoundSuffix}
+                                </span>
+                              )}
+                            </span>
+                          )}
+                          {/* Grounding check detail */}
+                          {step.step === 'grounding_check' && step.status === 'done' && (
+                            <span style={{ marginLeft: 6, color: step.sufficient ? 'var(--success-fg, #16a34a)' : 'var(--warning, #ca8a04)' }}>
+                              → {step.sufficient ? t.ragchat.groundingSufficient : t.ragchat.groundingInsufficient}
+                              {step.confidence !== undefined && ` (${Math.round(step.confidence * 100)}%)`}
+                            </span>
+                          )}
+                        </span>
+                      </div>
+                    );
+                  })}
+                </div>
+              )}
+            </div>
+          )}
+
+          {/* Messages area — accepts drag-and-drop document context attachment */}
+          <div
+            className="messages"
+            onDragOver={e => { e.preventDefault(); e.dataTransfer.dropEffect = 'copy'; }}
+            onDrop={handleFileDrop}
+          >
             {messages.map(msg => (
               <div key={msg.id} className={`message msg-${msg.role}`}>
                 {msg.role === 'assistant' && <div className="msg-avatar">AI</div>}
@@ -281,19 +623,130 @@ export function RagChatPage() {
                 </button>
               ))}
             </div>
+            {/* P0-1: Agentic mode toggle */}
+            <div style={{ display: 'flex', alignItems: 'center', gap: 8, marginBottom: 4 }}>
+              <label style={{
+                display: 'flex', alignItems: 'center', gap: 5,
+                fontSize: 12, color: agenticMode ? 'var(--accent, #6366f1)' : 'var(--muted)',
+                cursor: 'pointer', userSelect: 'none',
+              }}>
+                <input
+                  type="checkbox"
+                  checked={agenticMode}
+                  onChange={e => {
+                    setAgenticMode(e.target.checked);
+                    setThinkingSteps([]);
+                  }}
+                  style={{ cursor: 'pointer', accentColor: 'var(--accent, #6366f1)' }}
+                />
+                <Zap size={11} />
+                <span>{t.ragchat.agenticMode}</span>
+              </label>
+              {agenticMode && (
+                <span style={{ fontSize: 11, color: 'var(--muted)' }}>
+                  {t.ragchat.agenticModeHint}
+                </span>
+              )}
+            </div>
+
+            {/* ── Document context badge ── */}
+            {docContext && (
+              <div style={{
+                display: 'flex', alignItems: 'center', gap: 8,
+                padding: '6px 10px', marginBottom: 6,
+                background: docContext.status === 'error'
+                  ? 'rgba(220,38,38,0.06)'
+                  : docContext.status === 'ready'
+                  ? 'rgba(34,197,94,0.06)'
+                  : 'rgba(99,102,241,0.06)',
+                border: `1px solid ${
+                  docContext.status === 'error' ? 'rgba(220,38,38,0.3)'
+                  : docContext.status === 'ready' ? 'rgba(34,197,94,0.3)'
+                  : 'rgba(99,102,241,0.3)'
+                }`,
+                borderRadius: 8, fontSize: 12,
+              }}>
+                {docContext.status === 'extracting' && (
+                  <span style={{ animation: 'spin 1s linear infinite', display: 'inline-block', color: 'var(--accent,#6366f1)' }}>⟳</span>
+                )}
+                {docContext.status === 'ready' && <FileText size={13} color="#16a34a" />}
+                {docContext.status === 'error' && <AlertCircle size={13} color="#dc2626" />}
+
+                <span style={{
+                  fontWeight: 600, fontSize: 11,
+                  color: docContext.status === 'error' ? '#dc2626'
+                    : docContext.status === 'ready' ? '#16a34a'
+                    : 'var(--accent,#6366f1)',
+                  flexShrink: 0,
+                }}>
+                  {t.ragchat.attachContextBadge}
+                </span>
+
+                <span style={{ flex: 1, overflow: 'hidden', textOverflow: 'ellipsis', whiteSpace: 'nowrap', color: 'var(--fg)' }}
+                  title={docContext.filename}>
+                  {docContext.filename}
+                </span>
+
+                {docContext.status === 'ready' && (
+                  <span style={{ fontSize: 10, color: 'var(--muted)', flexShrink: 0 }}>
+                    {(docContext.charCount / 1000).toFixed(1)}k chars
+                    {docContext.truncated ? ` · ${t.ragchat.attachTruncated}` : ''}
+                  </span>
+                )}
+                {docContext.status === 'extracting' && (
+                  <span style={{ fontSize: 11, color: 'var(--accent,#6366f1)', flexShrink: 0 }}>
+                    {t.ragchat.attachExtracting}
+                  </span>
+                )}
+                {docContext.status === 'error' && (
+                  <span style={{ fontSize: 11, color: '#dc2626', flexShrink: 0 }} title={docContext.errorMsg}>
+                    {t.ragchat.attachError}
+                  </span>
+                )}
+
+                {/* Clear button */}
+                <button
+                  onClick={() => setDocContext(null)}
+                  style={{ background: 'none', border: 'none', cursor: 'pointer', padding: '2px 4px', color: 'var(--muted)', display: 'flex', alignItems: 'center', gap: 2, fontSize: 11, flexShrink: 0 }}
+                  title={t.ragchat.attachClearLabel}
+                >
+                  <X size={11} /> {t.ragchat.attachClearLabel}
+                </button>
+              </div>
+            )}
+
+            {/* Hidden file input */}
+            <input
+              ref={fileInputRef}
+              type="file"
+              accept={t.ragchat.attachAccept}
+              style={{ display: 'none' }}
+              onChange={handleFileInputChange}
+            />
+
             <div className="composer-row">
+              {/* Paperclip button — replaces attached doc when clicked again */}
+              <button
+                className="btn icon-btn"
+                onClick={() => fileInputRef.current?.click()}
+                disabled={streaming || docContext?.status === 'extracting'}
+                title={t.ragchat.attachBtn}
+                style={{ flexShrink: 0, padding: '8px', color: docContext?.status === 'ready' ? 'var(--accent, #6366f1)' : undefined }}
+              >
+                <Paperclip size={15} />
+              </button>
               <textarea
                 className="composer-input"
                 placeholder={t.ragchat.inputPlaceholder}
                 value={inputDraft}
                 onChange={e => setRagState(s => ({ ...s, inputDraft: e.target.value }))}
-                onKeyDown={e => { if (e.key === 'Enter' && !e.shiftKey) { e.preventDefault(); send(); } }}
+                onKeyDown={e => { if (e.key === 'Enter' && !e.shiftKey) { e.preventDefault(); void send(); } }}
                 rows={2}
               />
               <button
                 className="btn primary"
-                onClick={() => send()}
-                disabled={!inputDraft.trim() || streaming}
+                onClick={() => void send()}
+                disabled={!inputDraft.trim() || streaming || docContext?.status === 'extracting'}
               >
                 <Send size={14} />
               </button>