Upload 5 files

.gitattributes

@@ -35,3 +35,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
 faiss_hnsw/vector_search.index filter=lfs diff=lfs merge=lfs -text
 qdrant_db/collection/articles/storage.sqlite filter=lfs diff=lfs merge=lfs -text
+faiss_hnsw/index.faiss filter=lfs diff=lfs merge=lfs -text

app.py

@@ -7,6 +7,8 @@ import gradio as gr
 from llama_cpp import Llama
 from huggingface_hub import hf_hub_download
 from langchain_huggingface import HuggingFaceEmbeddings
+# Qdrant filter models
+from qdrant_client.http.models import Filter, FieldCondition, MatchValue
 
 # ====================== LOGGING ======================
 logging.basicConfig(level=logging.INFO, format="%(levelname)s | %(message)s")
@@ -128,51 +130,78 @@ def _load_llm() -> Llama:
 llm = _load_llm()
 
 # ====================== RAG CACHE ======================
-vectorstores: dict = {}
 
+# ====================== VECTOR STORES ======================
+vectorstores: dict = {}
 
 def get_vectorstore(backend_name: str):
     if backend_name in vectorstores:
         return vectorstores[backend_name]
+
     try:
         embeddings = HuggingFaceEmbeddings(
             model_name="BAAI/bge-small-en-v1.5",
             encode_kwargs={"normalize_embeddings": True},
         )
+
         if backend_name == "FAISS":
+            # Modern recommended import (still under langchain-community)
             from langchain_community.vectorstores import FAISS
-            vs = FAISS.load_local(FAISS_PATH, embeddings, allow_dangerous_deserialization=True)
+            vs = FAISS.load_local(
+                FAISS_PATH, 
+                embeddings, 
+                allow_dangerous_deserialization=True
+            )
+            logger.info("FAISS index loaded from %s", FAISS_PATH)
+
         elif backend_name == "Qdrant":
-            from langchain_community.vectorstores import Qdrant
-            vs = Qdrant(path=QDRANT_PATH, collection_name=QDRANT_COLLECTION, embeddings=embeddings)
+            # Modern Qdrant integration
+            from langchain_qdrant import QdrantVectorStore
+            from qdrant_client import QdrantClient
+
+            client = QdrantClient(
+                path=QDRANT_PATH,      # path to your qdrant_db folder
+                timeout=60,
+            )
+
+            vs = QdrantVectorStore(
+                client=client,
+                collection_name=QDRANT_COLLECTION,
+                embedding=embeddings,
+            )
+            logger.info("Qdrant collection '%s' loaded from %s", 
+                       QDRANT_COLLECTION, QDRANT_PATH)
+
         else:
+            # fallback to FAISS
             from langchain_community.vectorstores import FAISS
-            vs = FAISS.load_local(FAISS_PATH, embeddings, allow_dangerous_deserialization=True)
+            vs = FAISS.load_local(
+                FAISS_PATH, 
+                embeddings, 
+                allow_dangerous_deserialization=True
+            )
+
         vectorstores[backend_name] = vs
         logger.info("Vector store '%s' loaded successfully", backend_name)
         return vs
+
     except Exception as exc:
         logger.error("Failed to load vector store '%s': %s", backend_name, exc)
+        import traceback
+        logger.error(traceback.format_exc())
         return None
 
-
 def _rag_search(vs, query: str, k: int, article_filter: str, category_filter: str):
     """
     Similarity search with optional metadata filtering.
-
-    FAISS does not support dict-based server-side metadata filtering reliably
-    across langchain versions: it either ignores the filter silently or raises.
-    We therefore fetch a generous candidate pool and post-filter in Python.
-
-    Qdrant supports native dict filtering, so we pass it directly.
     """
     want_title    = None if article_filter  in (None, "", "All articles in category") else article_filter
     want_category = None if category_filter in (None, "", "All categories")           else category_filter
 
-    backend_type = type(vs).__name__   # "FAISS" or "Qdrant"
+    backend_type = type(vs).__name__
 
-    if backend_type == "FAISS":
-        # Fetch a large pool, then filter in Python.
+    if "FAISS" in backend_type:
+        # FAISS: post-filtering (unchanged)
         pool_size = min(k * 10, 80)
         pool = vs.similarity_search(query, k=pool_size)
 
@@ -188,34 +217,59 @@ def _rag_search(vs, query: str, k: int, article_filter: str, category_filter: st
                 break
 
         if not filtered and (want_title or want_category):
-            # Nothing matched the filter — warn and fall back to unfiltered results.
             logger.warning(
-                "FAISS post-filter (title=%r, cat=%r) matched 0 docs — "
-                "returning unfiltered top-%d",
-                want_title, want_category, k,
+                "FAISS post-filter (title=%r, cat=%r) matched 0 docs — returning unfiltered top-%d",
+                want_title, want_category, k
             )
             return pool[:k]
 
         logger.info(
             "FAISS post-filter (title=%r, cat=%r) → %d/%d docs kept",
-            want_title, want_category, len(filtered), len(pool),
+            want_title, want_category, len(filtered), len(pool)
         )
         return filtered
 
     else:
-        # Qdrant: use its native metadata filter dict.
-        filter_dict = None
+        # === QDRANT - FIXED METADATA FILTER ===
+        from qdrant_client.http.models import Filter, FieldCondition, MatchValue
+
+        conditions = []
+
         if want_title:
-            filter_dict = {"article_title": want_title}
+            conditions.append(
+                FieldCondition(
+                    key="metadata.article_title",      # ← Fixed: metadata. prefix
+                    match=MatchValue(value=want_title)
+                )
+            )
         elif want_category:
-            filter_dict = {"article_category": want_category}
+            conditions.append(
+                FieldCondition(
+                    key="metadata.article_category",   # ← Fixed: metadata. prefix
+                    match=MatchValue(value=want_category)
+                )
+            )
 
-        docs = vs.similarity_search(query, k=k, filter=filter_dict)
-        logger.info(
-            "Qdrant search (filter=%r) → %d docs", filter_dict, len(docs)
-        )
-        return docs
+        filter_dict = Filter(must=conditions) if conditions else None
+
+        try:
+            docs = vs.similarity_search(
+                query, 
+                k=k, 
+                filter=filter_dict
+            )
+            logger.info(
+                "Qdrant search (filter=%s) → %d docs", 
+                "title" if want_title else "category" if want_category else "none", 
+                len(docs)
+            )
+            return docs
 
+        except Exception as e:
+            logger.error("Qdrant search failed with filter: %s", e)
+            # Fallback: search without filter
+            logger.warning("Falling back to unfiltered Qdrant search")
+            return vs.similarity_search(query, k=k)
 
 # ====================== SYSTEM PROMPT ======================
 SYSTEM_PROMPT = """You are the reference expert for the articles contained in the training \

faiss_hnsw/index.faiss

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea56fb5d03cb5c7d4b8b243f81b383fc613927a7dc98956db3e506f1f623ff28
+size 553005

faiss_hnsw/index.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d5924462b6f8b8b7ce92dac44cacb29a8f61f6de0bc1e02f42831585db5e06e
+size 72893

requirements.txt

@@ -3,6 +3,7 @@ huggingface_hub
 llama-cpp-python @ https://huggingface.co/robertolofaro/libraries_prebuilt/resolve/main/llama_cpp_python-0.3.23-py3-none-linux_x86_64.whl
 langchain_huggingface
 langchain-community
+langchain-qdrant
 chromadb
 faiss-cpu
 qdrant-client