Add LLM token

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)
+