feat: implement basic RAG with pinecone

app.py

@@ -1,109 +1,83 @@
+import os
+import json
+
 import gradio as gr
 from openai import OpenAI
-import os
+from pinecone import Pinecone
 from dotenv import load_dotenv
-from datasets import load_dataset
-from sklearn.metrics.pairwise import cosine_similarity
-from functools import lru_cache
-import numpy as np
-
 
 load_dotenv(verbose=True)
 
-# Create OpenAI client (make sure OPENAI_API_KEY is set in your environment)
-client = OpenAI(
-    api_key=os.getenv("OPENAI_API_KEY"),
-)
-
-HF_DATASET = "carching/cs-data"
+PINECONE_API_KEY = os.getenv("PINECONE_API_KEY")
+INDEX_NAME = "whatsapp-history-1"
 EMBED_MODEL = "text-embedding-3-small"
 TOP_K = 5
 
-# ---- Load dataset ----
-def load_data():
-    ds = load_dataset(HF_DATASET)
-    messages = ds["train"]["message"]
-    senders = ds["train"]["sender"]
-    users = ds["train"]["user"]
-    return messages, senders, users
+client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
+pc = Pinecone(api_key=PINECONE_API_KEY)
+index = pc.Index(INDEX_NAME)
 
-messages, senders, users = load_data()
-
-# ---- Embedding helper ----
-@lru_cache(maxsize=None)
-def get_embedding(text):
-    resp = client.embeddings.create(
+def retrieve_context(query):
+    response = client.embeddings.create(
         model=EMBED_MODEL,
-        input=text
+        input=query
     )
-    return np.array(resp.data[0].embedding)
-
-# ---- Precompute dataset embeddings ----
-message_embeddings = [get_embedding(msg) for msg in messages]
-
-# ---- Retrieval ----
-def retrieve_context(user_query):
-    query_emb = get_embedding(user_query).reshape(1, -1)
-    all_embs = np.stack(message_embeddings)
-    sims = cosine_similarity(query_emb, all_embs)[0]
-    top_idx = np.argsort(sims)[-TOP_K:][::-1]
-    return [(messages[i], senders[i], sims[i]) for i in top_idx]
-
-# ---- Chatbot function ----
-def respond(message, history, system_message, max_tokens, temperature, top_p):
-    # Retrieve historical WhatsApp context
-    context_rows = retrieve_context(message)
-    context_text = "\n".join([f"{role}: {msg}" for msg, role, _ in context_rows])
-
-    # Build system prompt with context
-    full_system_message = (
-        f"{system_message}\n\n"
-        f"Use the following historical conversation context if relevant:\n\n"
-        f"{context_text}"
+    query_emb = response.data[0].embedding
+
+    # Use keyword argument 'vector' for the query
+    result = index.query(vector=query_emb, top_k=TOP_K, include_metadata=True)
+
+    contexts = []
+    for match in result.matches:
+        meta = match.metadata
+        contexts.append(f"{meta['sender']}: {meta['message']}")
+
+    return "\n".join(contexts)
+
+def respond(message, chat_history_json):
+    chat_history = json.loads(chat_history_json)
+
+    context = retrieve_context(message)
+
+    system_prompt = (
+        "You are a helpful assistant for carching. Use the following past conversation data on whatsapp "
+        "to answer the user's question if relevant:\n\n" + context
     )
 
-    # Convert Gradio history into OpenAI format
-    chat_messages = [{"role": "system", "content": full_system_message}]
-    for user_msg, assistant_msg in history:
-        if user_msg:
-            chat_messages.append({"role": "user", "content": user_msg})
-        if assistant_msg:
-            chat_messages.append({"role": "assistant", "content": assistant_msg})
-    chat_messages.append({"role": "user", "content": message})
-
-    # Stream responses from OpenAI
-    response_text = ""
-    stream = client.chat.completions.create(
-        model="gpt-4o-mini",
-        messages=chat_messages,
-        max_tokens=max_tokens,
-        temperature=temperature,
-        top_p=top_p,
-        stream=True
+    messages = [{"role": "system", "content": system_prompt}]
+    messages.extend(chat_history or [])
+    messages.append({"role": "user", "content": message})
+
+    response = client.chat.completions.create(
+        model="gpt-5-mini",
+        messages=messages,
+        temperature=0.7
     )
 
-    for chunk in stream:
-        if chunk.choices[0].delta.content:
-            token = chunk.choices[0].delta.content
-            response_text += token
-            yield response_text
-
-# ---- Gradio Interface ----
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a helpful support assistant.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)"
-        ),
-    ],
-)
-
-if __name__ == "__main__":
-    demo.launch()
+    bot_reply = response.choices[0].message.content
+
+    chat_history.append({"role": "user", "content": message})
+    chat_history.append({"role": "assistant", "content": bot_reply})
+
+    # Return the history directly and the serialized state
+    return chat_history, json.dumps(chat_history)
+
+with gr.Blocks() as demo:
+    gr.Markdown("# Customer Support Chatbot")
+    # Set chatbot type to 'messages' to resolve the warning
+    chatbot = gr.Chatbot(type="messages")
+    msg = gr.Textbox(placeholder="Ask a question...", show_label=False)
+    state = gr.State(json.dumps([]))
+
+    with gr.Row():
+        submit_btn = gr.Button("Send")
+
+    def submit_message(msg, state):
+        # Clear the textbox after submission
+        return "", *respond(msg, state)
+
+    # Bind submit button AND hitting enter in textbox
+    submit_btn.click(fn=submit_message, inputs=[msg, state], outputs=[msg, chatbot, state])
+    msg.submit(fn=submit_message, inputs=[msg, state], outputs=[msg, chatbot, state])
+
+demo.launch()

requirements.txt

@@ -4,4 +4,5 @@ openai
 dotenv
 datasets
 scikit-learn
-numpy
+numpy
+pinecone