Add LLM token

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]: