Add LLM token

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()