Create qa.py

src/qa.py

@@ -0,0 +1,63 @@
+import os
+from sentence_transformers import SentenceTransformer
+from transformers import pipeline
+from vectorstore import search_faiss
+
+print("✅ qa.py loaded from:", __file__)
+
+# Always redirect Hugging Face caches to /tmp
+CACHE_DIR = "/tmp/huggingface"
+os.environ["HF_HOME"] = CACHE_DIR
+os.environ["TRANSFORMERS_CACHE"] = CACHE_DIR
+os.environ["HF_DATASETS_CACHE"] = CACHE_DIR
+
+# ----------------------------
+# Embedding model for queries
+# ----------------------------
+_query_model = SentenceTransformer(
+    "sentence-transformers/all-MiniLM-L6-v2",
+    cache_folder=CACHE_DIR
+)
+
+# ----------------------------
+# LLM for answers
+# ----------------------------
+MODEL_NAME = "google/flan-t5-small"
+MODEL_PATH = os.path.join(CACHE_DIR, MODEL_NAME)
+
+if not os.path.exists(MODEL_PATH):
+    print(f"⬇️ Downloading {MODEL_NAME} to {MODEL_PATH}")
+    _answer_model = pipeline(
+        "text2text-generation",
+        model=MODEL_NAME,
+        cache_dir=CACHE_DIR
+    )
+    # Save pipeline model locally
+    _answer_model.model.save_pretrained(MODEL_PATH)
+    _answer_model.tokenizer.save_pretrained(MODEL_PATH)
+else:
+    print(f"✅ Loading {MODEL_NAME} from {MODEL_PATH}")
+    _answer_model = pipeline(
+        "text2text-generation",
+        model=MODEL_PATH,
+        cache_dir=CACHE_DIR
+    )
+
+# ----------------------------
+# Functions
+# ----------------------------
+def retrieve_chunks(query, index, chunks, top_k=3):
+    q_emb = _query_model.encode([query], convert_to_numpy=True)[0]
+    return search_faiss(q_emb, index, chunks, top_k)
+
+def generate_answer(query, retrieved_chunks):
+    if not retrieved_chunks:
+        return "Sorry, I could not find relevant information."
+
+    context = " ".join(retrieved_chunks)
+    prompt = (
+        "You are an assistant. Use the context to answer the question clearly.\n"
+        f"Context:\n{context}\n\nQuestion:\n{query}\n\nAnswer:"
+    )
+    result = _answer_model(prompt, max_length=300, do_sample=False)
+    return result[0]["generated_text"].strip()