Update app.py

app.py

@@ -455,27 +455,18 @@ def get_response_from_pdf(query, model, selected_docs, num_calls=3, temperature=
     if os.path.exists("faiss_database"):
         logging.info("Loading FAISS database")
         database = FAISS.load_local("faiss_database", embed, allow_dangerous_deserialization=True)
-        
-        # Inspect FAISS database
-        logging.info(f"FAISS database size: {len(database.docstore._dict)}")
-        for doc_id, doc in database.docstore._dict.items():
-            logging.info(f"Document ID: {doc_id}, Source: {doc.metadata.get('source', 'Unknown')}")
     else:
         logging.warning("No FAISS database found")
         yield "No documents available. Please upload PDF documents to answer questions."
         return
 
-    retriever = database.as_retriever(search_kwargs={"k": 20})  # Increased k to 20
+    retriever = database.as_retriever(search_kwargs={"k": 20})  # Increased k to 10
     logging.info(f"Retrieving relevant documents for query: {query}")
     relevant_docs = retriever.get_relevant_documents(query)
     logging.info(f"Number of relevant documents retrieved: {len(relevant_docs)}")
     
-    # Log details of retrieved documents
-    for i, doc in enumerate(relevant_docs):
-        logging.info(f"Retrieved document {i+1}: Source: {doc.metadata.get('source', 'Unknown')}, Content preview: {doc.page_content[:100]}...")
-
     # Filter relevant_docs based on selected documents
-    filtered_docs = [doc for doc in relevant_docs if doc.metadata.get("source") in selected_docs]
+    filtered_docs = [doc for doc in relevant_docs if doc.metadata["source"] in selected_docs]
     logging.info(f"Number of filtered documents: {len(filtered_docs)}")
     
     if not filtered_docs:
@@ -483,28 +474,28 @@ def get_response_from_pdf(query, model, selected_docs, num_calls=3, temperature=
         yield "No relevant information found in the selected documents. Please try selecting different documents or rephrasing your query."
         return
 
-    # Implement a custom chunking strategy
-    def custom_chunk(text, chunk_size=1000, overlap=200):
-        chunks = []
-        start = 0
-        while start < len(text):
-            end = start + chunk_size
-            chunk = text[start:end]
-            chunks.append(chunk)
-            start = end - overlap
-        return chunks
+    for doc in filtered_docs:
+        logging.info(f"Document source: {doc.metadata['source']}")
+        logging.info(f"Document content preview: {doc.page_content[:100]}...")
 
+    # Implement a sliding window approach for context
+    max_context_length = 4000  # Adjust based on your model's capacity
     context_chunks = []
+    current_chunk = ""
     for doc in filtered_docs:
-        doc_chunks = custom_chunk(doc.page_content)
-        context_chunks.extend(doc_chunks)
+        if len(current_chunk) + len(doc.page_content) > max_context_length:
+            context_chunks.append(current_chunk)
+            current_chunk = doc.page_content
+        else:
+            current_chunk += "\n" + doc.page_content
+    if current_chunk:
+        context_chunks.append(current_chunk)
 
-    logging.info(f"Number of context chunks after custom chunking: {len(context_chunks)}")
+    logging.info(f"Number of context chunks: {len(context_chunks)}")
 
     for i, context_str in enumerate(context_chunks):
         logging.info(f"Processing context chunk {i+1}/{len(context_chunks)}")
         logging.info(f"Context chunk length: {len(context_str)}")
-        logging.info(f"Context chunk preview: {context_str[:100]}...")
 
         if model == "@cf/meta/llama-3.1-8b-instruct":
             logging.info("Using Cloudflare API")
@@ -523,7 +514,7 @@ Write a detailed and complete response that answers the following user question:
                 logging.info(f"API call {j+1}/{num_calls}")
                 for message in client.chat_completion(
                     messages=[{"role": "user", "content": prompt}],
-                    max_tokens=2000,  # Reduced max_tokens to avoid potential errors
+                    max_tokens=10000,
                     temperature=temperature,
                     stream=True,
                 ):