Add knowledge retrieval configuration to Assistant model and related components

- Introduce new fields for knowledge retrieval configuration in AssistantConfig and Assistant models, including mode, top_n, and score_threshold.
- Implement KnowledgeRetrievalConfig schema with validation for top_n.
- Update backend services and routes to handle knowledge retrieval settings.
- Enhance frontend components to support knowledge retrieval configuration, including a new dialog for advanced settings.
- Add tests for knowledge retrieval configuration validation and description generation.

backend/services/knowledge.py

@@ -203,7 +203,13 @@ async def recover_interrupted_documents() -> None:
         await session.commit()
 
 
-async def search(session: AsyncSession, kb_id: str, query: str, top_k: int | None = None) -> list[dict]:
+async def search(
+    session: AsyncSession,
+    kb_id: str,
+    query: str,
+    top_k: int | None = None,
+    score_threshold: float = 0.0,
+) -> list[dict]:
     kb = await session.get(KnowledgeBase, kb_id)
     if not kb:
         return []
@@ -217,13 +223,20 @@ async def search(session: AsyncSession, kb_id: str, query: str, top_k: int | Non
         return []
     query_embedding = (await _embed(session, kb, [query]))[0]
     distance = KnowledgeChunk.embedding.cosine_distance(query_embedding)
-    rows = (await session.execute(
+    statement = (
         select(KnowledgeChunk, KnowledgeDocument.name, distance.label("distance"))
         .join(KnowledgeDocument, KnowledgeDocument.id == KnowledgeChunk.document_id)
-        .where(KnowledgeChunk.knowledge_base_id == kb_id, KnowledgeDocument.status == "ready")
+        .where(
+            KnowledgeChunk.knowledge_base_id == kb_id,
+            KnowledgeDocument.status == "ready",
+            distance <= 1.0 - score_threshold,
+        )
         .order_by(distance)
-        .limit(top_k or settings.KNOWLEDGE_TOP_K)
-    )).all()
+    )
+    effective_top_k = settings.KNOWLEDGE_TOP_K if top_k is None else top_k
+    if effective_top_k != -1:
+        statement = statement.limit(effective_top_k)
+    rows = (await session.execute(statement)).all()
     return [
         {"content": chunk.content, "document": name, "score": round(max(0.0, 1.0 - float(dist)), 4)}
         for chunk, name, dist in rows