Update app.py

app.py

@@ -88,7 +88,15 @@ def generate_qa_chain_with_citations(repo_id, llm):
     embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")
     
     # Download files
-    faiss_path = hf_hub_download(repo_id=repo_id, filename="index.faiss", repo_type="dataset")
+    # In generate_qa_chain_with_citations(), replace:
+    faiss_path = hf_hub_download(
+        repo_id=repo_id, 
+        filename="index.faiss", 
+        repo_type="dataset",
+        cache_dir="/tmp/hf_cache"  # Dedicated cache
+    )
+
+    #faiss_path = hf_hub_download(repo_id=repo_id, filename="index.faiss", repo_type="dataset")
     pkl_path = hf_hub_download(repo_id=repo_id, filename="index.pkl", repo_type="dataset")
     metadata_path = hf_hub_download(repo_id=repo_id, filename="metadata.json", repo_type="dataset")
     
@@ -154,6 +162,23 @@ def format_citations_with_links(sources, uploaded_files):
         citations_html.append(citation_html)
     
     return "".join(citations_html)
+#=========================================
+
+from langchain_huggingface import HuggingFacePipeline
+from transformers import pipeline
+
+LLM_CACHE = None  # Global cache
+
+def get_cached_llm():
+    global LLM_CACHE
+    if LLM_CACHE is None:
+        LLM_CACHE = HuggingFacePipeline.from_model_id(
+            model_id="distilgpt2",  # Smallest, fastest
+            task="text-generation",
+            device_map="cpu",
+            pipeline_kwargs={"max_new_tokens": 100}
+        )
+    return LLM_CACHE
 
 # ========================================
 # Creating the llm with model
@@ -177,7 +202,8 @@ def create_llm_pipeline():
 # ========================================
 def rag_query_with_citations(question, repo_id, history=[], uploaded_files=[]):
     try:
-        llm = create_llm_pipeline()
+        #llm = create_llm_pipeline()
+        llm = get_cached_llm()  # Single creation
         qa_chain, metadata_path = generate_qa_chain_with_citations(repo_id, llm)
         
         result = qa_chain.invoke({"query": question})
@@ -232,6 +258,7 @@ with gr.Blocks(title="NRL Chat for Commercial procurement", theme=gr.themes.Soft
                     if f:
                         with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp:
                             tmp.write(f.read())
+                            tmp.close()  # Explicit close
                             file_dict[f.name] = tmp.name
                 return file_dict