Update app.py

app.py

@@ -1,83 +1,65 @@
-import gradio as gr
-import os
-import json
 import faiss
 import numpy as np
 import torch
 from sentence_transformers import SentenceTransformer
-from huggingface_hub import InferenceClient, hf_hub_download
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import gradio as gr
+from huggingface_hub import hf_hub_download
 
 # 🔹 Hugging Face Credentials
 HF_REPO = "Futuresony/future_ai_12_10_2024.gguf"
-HF_TOKEN = os.getenv('HUGGINGFACEHUB_API_TOKEN')  # Ensure this is set in your environment
+HF_TOKEN = os.getenv('HUGGINGFACEHUB_API_TOKEN')
 
-# 🔹 FAISS Index Path
+# 🔹 Paths
 FAISS_PATH = "asa_faiss.index"
+DATASET_PATH = "responses.txt"
 
-# 🔹 Load Sentence Transformer for Embeddings
-embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
-
-# 🔹 Load FAISS Index from Hugging Face
-faiss_local_path = hf_hub_download(HF_REPO, "asa_faiss.index", token=HF_TOKEN)
+# 🔹 Load FAISS Index
+faiss_local_path = hf_hub_download(HF_REPO, FAISS_PATH, token=HF_TOKEN)
 faiss_index = faiss.read_index(faiss_local_path)
+print("✅ FAISS index loaded successfully!")
 
-# 🔹 Initialize Hugging Face Model Client
-client = InferenceClient(model=HF_REPO, token=HF_TOKEN)
+# 🔹 Load Dataset Responses
+with open(DATASET_PATH, "r", encoding="utf-8") as f:
+    dataset = f.readlines()
+print("✅ Responses dataset loaded!")
 
-# 🔹 Retrieve Relevant FAISS Data
-def retrieve_faiss_knowledge(user_query, top_k=3):
-    query_embedding = embedder.encode([user_query], convert_to_tensor=True).cpu().numpy()
-    distances, indices = faiss_index.search(query_embedding, top_k)
+# 🔹 Load Model & Tokenizer Correctly
+tokenizer = AutoTokenizer.from_pretrained(HF_REPO, token=HF_TOKEN)
+model = AutoModelForCausalLM.from_pretrained(HF_REPO, token=HF_TOKEN)
 
-    retrieved_texts = []
-    for idx in indices[0]:  # Extract top_k results
-        if idx != -1:  # Ensure valid index
-            retrieved_texts.append(f"Example {idx}: (Extracted FAISS Data)")
-
-    return "\n".join(retrieved_texts) if retrieved_texts else "**No relevant FAISS data found.**"
-
-# 🔹 Chatbot Response Function (Forcing FAISS Context)
-def respond(message, history, system_message, max_tokens, temperature, top_p):
-    faiss_context = retrieve_faiss_knowledge(message)
+# 🔹 Load Sentence Transformer (Ensures proper FAISS embedding match)
+embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
 
-    # 🔥 Force the model to use FAISS
-    full_prompt = f"""### System Instruction:
-You MUST use the provided FAISS data to generate your response.
-If no FAISS data is found, return "I don't have enough information."
+# 🔹 Adjust FAISS Threshold
+THRESHOLD = 100  # Adjust based on actual FAISS distances
 
-### Retrieved FAISS Data:
-{faiss_context}
 
-### User Query:
-{message}
+def embed(text):
+    """Convert text into FAISS-compatible vector using the same method as training."""
+    return embedder.encode([text], convert_to_tensor=True).cpu().numpy()
 
-### Response:
-"""
 
-    response = client.text_generation(
-        full_prompt,
-        max_new_tokens=max_tokens,
-        temperature=temperature,
-        top_p=top_p,
-    )
+def chatbot_response(user_query):
+    """Fetch response from FAISS or generate with model if needed."""
+    query_vector = embed(user_query)  # Convert input to vector
+    D, I = faiss_index.search(query_vector, k=1)  # Search FAISS index
 
-    # ✅ Extract only the model-generated response
-    cleaned_response = response.split("### Response:")[-1].strip()
+    print(f"Closest FAISS match index: {I[0][0]}, Distance: {D[0][0]}")  # Debugging info
 
-    history.append((message, cleaned_response))  # ✅ Update chat history
+    if D[0][0] < THRESHOLD:  # If FAISS match is good
+        response = dataset[I[0][0]].strip()  # Retrieve FAISS response
+        print("✅ FAISS response used!")
+    else:
+        # 🔥 Fallback: Generate response with model
+        print("⚠️ FAISS match too weak, using model instead.")
+        inputs = tokenizer(user_query, return_tensors="pt")
+        outputs = model.generate(**inputs, max_new_tokens=150)
+        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
 
-    yield cleaned_response  # ✅ Output the response
+    return response
 
-# 🔹 Gradio Chat Interface
-demo = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a knowledge assistant that must use FAISS context.", label="System message"),
-        gr.Slider(minimum=1, maximum=250, value=128, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.9, step=0.1, label="Temperature"),
-        gr.Slider(minimum=0.1, maximum=1.0, value=0.99, step=0.01, label="Top-p (nucleus sampling)"),
-    ],
-)
 
-if __name__ == "__main__":
-    demo.launch()
+# 🔹 Gradio UI
+iface = gr.Interface(fn=chatbot_response, inputs="text", outputs="text", title="ASA Microfinance Chatbot")
+iface.launch()