Enhance knowledge base functionality and integrate S3 storage support

- Add new models for `KnowledgeDocument` and `KnowledgeChunk` to manage document ingestion and chunking.
- Implement S3-compatible storage integration for knowledge documents, allowing for file uploads and retrieval.
- Introduce API endpoints for managing knowledge bases and documents, including creation, deletion, and searching.
- Update frontend components to support knowledge base configuration and document management, improving user interaction.
- Enhance backend services for knowledge processing and retrieval, ensuring robust handling of document statuses and errors.

backend/services/pipecat/pipeline.py

@@ -27,6 +27,8 @@ from services.pipecat.service_factory import (
     create_stt,
     create_tts,
 )
+from db.session import SessionLocal
+from services.knowledge import search as search_knowledge
 
 from pipecat.adapters.schemas.function_schema import FunctionSchema
 from pipecat.adapters.schemas.tools_schema import ToolsSchema
@@ -36,6 +38,7 @@ from pipecat.frames.frames import (
     InterruptionFrame,
     LLMFullResponseEndFrame,
     LLMFullResponseStartFrame,
+    LLMContextFrame,
     LLMTextFrame,
     LLMMessagesAppendFrame,
     OutputTransportMessageUrgentFrame,
@@ -79,6 +82,12 @@ VISION_ANALYSIS_SYSTEM_PROMPT = (
     "你是一个视觉理解模型。请只根据图片内容和用户问题给出准确、简洁的中文观察结果。"
     "如果画面不足以判断,请明确说明不确定。"
 )
+KNOWLEDGE_TOOL_NAME = "search_knowledge_base"
+KNOWLEDGE_SYSTEM_HINT = (
+    "你已连接内部知识库。系统会在每轮用户问题前自动提供相关资料；"
+    "回答资料事实时只根据检索内容，资料不足要明确说明。"
+)
+KNOWLEDGE_CONTEXT_MARKER = "<!-- knowledge-context -->"
 
 
 def _require(value: str, label: str) -> str:
@@ -309,6 +318,54 @@ class ConversationHistoryProcessor(FrameProcessor):
             await self._recorder.record_transport_message(frame.message)
 
 
+class KnowledgeRetrievalProcessor(FrameProcessor):
+    """Retrieve before local LLM inference without changing Pipecat internals."""
+
+    def __init__(self, knowledge_base_id: str | None):
+        super().__init__()
+        self._knowledge_base_id = knowledge_base_id
+        self._last_signature = ""
+
+    async def process_frame(self, frame, direction: FrameDirection):
+        await super().process_frame(frame, direction)
+        if not self._knowledge_base_id or not isinstance(frame, LLMContextFrame):
+            await self.push_frame(frame, direction)
+            return
+
+        messages = frame.context.get_messages()
+        user_messages = [message for message in messages if message.get("role") == "user"]
+        if not user_messages:
+            await self.push_frame(frame, direction)
+            return
+        query = str(user_messages[-1].get("content") or "").strip()
+        signature = f"{len(user_messages)}:{query}"
+        if not query or signature == self._last_signature:
+            await self.push_frame(frame, direction)
+            return
+        self._last_signature = signature
+
+        try:
+            async with SessionLocal() as session:
+                results = await search_knowledge(session, self._knowledge_base_id, query)
+        except Exception as exc:
+            logger.warning(f"自动知识库检索失败: {exc}")
+            results = []
+
+        sources = "\n\n".join(
+            f"[{index + 1}] 来源：{item['document']}（相关度 {item['score']}）\n{item['content']}"
+            for index, item in enumerate(results)
+        ) or "未检索到相关资料。"
+        block = f"{KNOWLEDGE_CONTEXT_MARKER}\n当前问题的知识库检索结果：\n{sources}"
+        system_message = next((message for message in messages if message.get("role") == "system"), None)
+        if system_message is None:
+            messages.insert(0, {"role": "system", "content": block})
+        else:
+            content = str(system_message.get("content") or "")
+            base = content.split(KNOWLEDGE_CONTEXT_MARKER, 1)[0].rstrip()
+            system_message["content"] = f"{base}\n\n{block}" if base else block
+        await self.push_frame(frame, direction)
+
+
 class PassthroughLLMAssistantAggregator(LLMAssistantAggregator):
     """聚合 LLM 回复进上下文,同时继续把回复帧交给下游 TTS。"""
 
@@ -432,11 +489,12 @@ async def run_pipeline(
     call_end = CallEndCoordinator(queue_call_end)
 
     def with_vision_hint(text: str) -> str:
-        if not vision_enabled:
-            return text
-        if not text:
-            return VISION_SYSTEM_HINT
-        return f"{text}\n\n{VISION_SYSTEM_HINT}"
+        hints = []
+        if vision_enabled:
+            hints.append(VISION_SYSTEM_HINT)
+        if cfg.knowledge_base_id:
+            hints.append(KNOWLEDGE_SYSTEM_HINT)
+        return "\n\n".join(part for part in [text, *hints] if part)
 
     context = LLMContext(
         messages=[{"role": "system", "content": with_vision_hint(system_content)}]
@@ -460,6 +518,7 @@ async def run_pipeline(
     assistant_aggregator = PassthroughLLMAssistantAggregator(context)
     text_input = TextInputProcessor(should_ignore_input=lambda: call_end.ending)
     vision_capture = VisionCaptureProcessor()
+    knowledge_retrieval = KnowledgeRetrievalProcessor(cfg.knowledge_base_id)
     vision_native_mode = vision_enabled and _vision_uses_main_llm(cfg)
     vision_state: dict[str, str | None] = {"client_id": None}
     vision_schema = FunctionSchema(
@@ -476,6 +535,27 @@ async def run_pipeline(
         },
         required=["question"],
     )
+    knowledge_schema = FunctionSchema(
+        name=KNOWLEDGE_TOOL_NAME,
+        description="在当前助手绑定的知识库中检索与问题最相关的资料片段。",
+        properties={
+            "query": {"type": "string", "description": "用于检索的完整问题或关键词"}
+        },
+        required=["query"],
+    )
+
+    async def search_bound_knowledge(params: FunctionCallParams):
+        query = str(params.arguments.get("query") or "").strip()
+        if not query or not cfg.knowledge_base_id:
+            await params.result_callback({"status": "error", "message": "检索问题为空或未绑定知识库"})
+            return
+        try:
+            async with SessionLocal() as session:
+                results = await search_knowledge(session, cfg.knowledge_base_id, query)
+            await params.result_callback({"status": "ok", "results": results})
+        except Exception as exc:
+            logger.exception(f"知识库检索失败: {exc}")
+            await params.result_callback({"status": "error", "message": "知识库检索暂时不可用"})
 
     async def fetch_user_image(params: FunctionCallParams):
         question = str(params.arguments.get("question") or "请描述当前画面。")
@@ -538,11 +618,15 @@ async def run_pipeline(
 
     if vision_enabled:
         llm.register_function(VISION_TOOL_NAME, fetch_user_image)
+    if cfg.knowledge_base_id:
+        llm.register_function(KNOWLEDGE_TOOL_NAME, search_bound_knowledge)
 
     def set_visible_tools(schemas: list[FunctionSchema] | None = None) -> None:
         tools = list(schemas or [])
         if vision_enabled:
             tools.append(vision_schema)
+        if cfg.knowledge_base_id:
+            tools.append(knowledge_schema)
         if tools:
             context.set_tools(ToolsSchema(standard_tools=tools))
         else:
@@ -561,6 +645,7 @@ async def run_pipeline(
             text_input,
             stt,
             user_aggregator,
+            knowledge_retrieval,
             llm,
             # Aggregate the streamed LLM text before TTS. On interruption,
             # Pipecat commits the generated prefix immediately instead of