Upload app.py

app.py

@@ -5,7 +5,7 @@ import gradio as gr
 import torch
 from dotenv import load_dotenv
 from pinecone import Pinecone
-from langchain_huggingface import HuggingFaceEmbeddings
+from langchain.embeddings.huggingface import HuggingFaceEmbeddings
 
 # Detect GPU availability and set device
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
@@ -39,48 +39,74 @@ index = initialize_pinecone_index(index_name)
 # Initialize HuggingFace embedding model
 embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/msmarco-distilbert-base-v4")
 
+# Initialize chat history manually
+chat_history = []
+
+# Helper function to recursively flatten any list to a string
+def flatten_to_string(data):
+    if isinstance(data, list):
+        return " ".join([flatten_to_string(item) for item in data])
+    if data is None:
+        return ""
+    return str(data)
+
 # Function to interact with Pinecone and OpenAI GPT-4
-def get_model_response(human_input, chat_history=None):
+def get_model_response(human_input):
     try:
-        # Fetching and embedding query (move tensors to GPU)
+        # Embed the query
         query_embedding = torch.tensor(embedding_model.embed_query(human_input)).to(device)
-
-        # Convert NumPy array to list before passing it to Pinecone or any API that requires JSON-serializable data
         query_embedding = query_embedding.cpu().numpy().tolist()
 
-        # Query Pinecone index using CPU or GPU for fast vector search (if supported)
+        # Query Pinecone index
         search_results = index.query(vector=query_embedding, top_k=2, include_metadata=True)
 
         context_list, images = [], []
         for ind, result in enumerate(search_results['matches']):
-            document_content = result.get('metadata', {}).get('content', 'No content found')
-            image_url = result.get('metadata', {}).get('image_path', None)
-            figure_desc = result.get('metadata', {}).get('figure_description', '')
-            
+            document_content = flatten_to_string(result.get('metadata', {}).get('content', 'No content found'))
+            image_url = flatten_to_string(result.get('metadata', {}).get('image_path', None))
+            figure_desc = flatten_to_string(result.get('metadata', {}).get('figure_description', ''))
+
             context_list.append(f"Document {ind+1}: {document_content}")
-            
             if image_url and figure_desc:
                 images.append((figure_desc, image_url))
-        
+
         context_string = '\n\n'.join(context_list)
-        messages = [{"role": "system", "content": "You are a helpful assistant."},
-                    {"role": "user", "content": f"Here is some context:\n{context_string}\n\nUser's question: {human_input}"}]
-        
-        # Generate response using OpenAI GPT-3.5 Turbo for faster responses
+
+        # Add user message to chat history
+        chat_history.append({"role": "user", "content": human_input})
+
+        # Create messages for OpenAI's API
+        messages = [{"role": "system", "content": "You are a helpful assistant."}] + chat_history + [
+            {"role": "system", "content": f"Here is some context:\n{context_string}"},
+            {"role": "user", "content": human_input}
+        ]
+
+        # Validate messages before sending to OpenAI
+        for message in messages:
+            if not isinstance(message, dict) or "role" not in message or "content" not in message:
+                raise ValueError(f"Invalid message format: {message}")
+
+        # Send the conversation to OpenAI's API
         response = openai.ChatCompletion.create(
             model="gpt-3.5-turbo", 
             messages=messages, 
             max_tokens=500, 
             temperature=0.5
         )
+
         output_text = response['choices'][0]['message']['content'].strip()
+
+        # Add assistant message to chat history
+        chat_history.append({"role": "assistant", "content": output_text})
+
         return output_text, images
+
     except Exception as e:
         return f"Error invoking model: {str(e)}", []
 
-# Function to format text and images for display
-def get_model_response_with_images(human_input, chat_history=None):
-    output_text, images = get_model_response(human_input, chat_history)
+# Function to format text and images for display and track conversation
+def get_model_response_with_images(human_input, history=None):
+    output_text, images = get_model_response(human_input)
     if images:
         image_output = "".join([f"\n\n**{figure_desc}**\n![{figure_desc}]({image_path})" for figure_desc, image_path in images])
         return output_text + image_output