Update api/server.py

api/server.py

@@ -391,7 +391,6 @@ def login(req: LoginReq):
     sess["name"] = name
     return {"ok": True, "user": {"name": name, "user_id": user_id}}
 
-
 @app.post("/api/chat")
 def chat(req: ChatReq):
     user_id = (req.user_id or "").strip()
@@ -412,6 +411,37 @@ def chat(req: ChatReq):
             "run_id": None,
         }
 
+    # ----------------------------
+    # RAG query normalization (short file-intent prompts)
+    # ----------------------------
+    def _looks_like_file_request(text: str) -> bool:
+        t = (text or "").strip().lower()
+        if not t:
+            return False
+        triggers = [
+            "read this",
+            "summarize",
+            "summary",
+            "can you see",
+            "see that file",
+            "see the file",
+            "that file",
+            "this file",
+            "the file",
+            "attached",
+            "attachment",
+            "upload",
+            "uploaded",
+            "document",
+            "pdf",
+            "ppt",
+            "slides",
+            "docx",
+            "analyze",
+            "explain this doc",
+        ]
+        return any(k in t for k in triggers)
+
     t0 = time.time()
     marks_ms: Dict[str, float] = {"start": 0.0}
 
@@ -424,10 +454,33 @@ def chat(req: ChatReq):
     sess["cognitive_state"] = update_cognitive_state_from_message(msg, sess["cognitive_state"])
     marks_ms["cognitive_update_done"] = (time.time() - t0) * 1000.0
 
-    if len(msg) < 20 and ("?" not in msg):
+    # ✅ Key fix:
+    # - DO NOT gate RAG purely by message length.
+    # - For very short generic messages (e.g. "hi"), skip to save latency.
+    # - For short file-intent messages (e.g. "Read this"), force a better retrieval query.
+    rag_context_text, rag_used_chunks = "", []
+    try:
+        chunks = sess.get("rag_chunks") or []
+        has_chunks = len(chunks) > 0
+
+        if has_chunks:
+            is_file_intent = _looks_like_file_request(msg)
+            is_too_short_generic = (len(msg) < 8) and (not is_file_intent)
+
+            if is_too_short_generic:
+                rag_context_text, rag_used_chunks = "", []
+            else:
+                retrieval_query = msg
+                if is_file_intent:
+                    # Make retrieval robust even when user message is vague.
+                    # Include doc_type to bias retrieval toward recently uploaded material.
+                    retrieval_query = f"uploaded document ({req.doc_type}) key content and relevant excerpts for: {msg}"
+
+                rag_context_text, rag_used_chunks = retrieve_relevant_chunks(retrieval_query, chunks)
+    except Exception as e:
+        print(f"[chat] rag retrieve error: {repr(e)}")
         rag_context_text, rag_used_chunks = "", []
-    else:
-        rag_context_text, rag_used_chunks = retrieve_relevant_chunks(msg, sess["rag_chunks"])
+
     marks_ms["rag_retrieve_done"] = (time.time() - t0) * 1000.0
 
     try:
@@ -496,7 +549,7 @@ def chat(req: ChatReq):
             "learning_mode": req.learning_mode,
             "doc_type": req.doc_type,
             "refs": refs,
-            "run_id": run_id,  # NEW: keep in dataset metadata for debugging
+            "run_id": run_id,
         }
     )
 
@@ -507,7 +560,7 @@ def chat(req: ChatReq):
         ),
         "refs": refs,
         "latency_ms": total_ms,
-        "run_id": run_id,  # NEW: FE attaches feedback to this run
+        "run_id": run_id,
     }