modified: blueprints/chat.py

modified:   services/rag_service.py

blueprints/chat.py

@@ -63,6 +63,36 @@ def ping_llm():
         return jsonify({"status": "error", "error": str(e), "url": url, "model": model, "embed_model": embed_model}), 502
 
 
+@chat_bp.route("/debug-rag", methods=["GET"])
+@login_required
+def debug_rag():
+    """Diagnose RAG: show collection size, run a test query, return raw distances."""
+    query = request.args.get("q", "test")
+    source_id = request.args.get("source_id", type=int)
+    source_type = request.args.get("source_type", "doc")
+    try:
+        collection = rag_service._get_collection()
+        total_chunks = collection.count()
+        source_ids = [source_id] if source_id else None
+        raw = collection.query(
+            query_texts=[query],
+            n_results=min(5, max(1, total_chunks)),
+            where=rag_service._build_where(current_user.id, source_ids, source_type if source_id else None),
+            include=["documents", "distances", "ids"],
+        )
+        hits = [
+            {"id": i, "distance": round(d, 4), "preview": t[:120]}
+            for i, d, t in zip(
+                (raw.get("ids") or [[]])[0],
+                (raw.get("distances") or [[]])[0],
+                (raw.get("documents") or [[]])[0],
+            )
+        ]
+        return jsonify({"total_chunks": total_chunks, "query": query, "hits": hits})
+    except Exception as e:
+        return jsonify({"error": str(e)}), 500
+
+
 # ── Main chat ────────────────────────────────────────────────────────────────
 
 @chat_bp.route("/sessions/<int:session_id>/ask", methods=["POST"])
@@ -108,6 +138,8 @@ def ask(session_id):
     except Exception as e:
         current_app.logger.warning(f"RAG lookup failed, continuing without context: {e}")
 
+    current_app.logger.info(f"RAG: found {len(chunks)} chunks for session {session_id}")
+
     # Build history (last 10 messages for context window)
     history = [
         {"role": m.role, "content": m.content}