Update app.py

app.py

@@ -1,5 +1,5 @@
-import gradio as gr
 
+import gradio as gr
 import torch
 import requests
 from transformers import pipeline
@@ -10,11 +10,18 @@ import dspy
 import json
 
 # === Load Models ===
+print("Loading zero-shot classifier...")
 classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
+
+print("Loading embedding model...")
 embedding_model = SentenceTransformer("intfloat/e5-large")
-qa_pipeline = pipeline("text-generation", model="WizardLM/WizardMath-7B-V1.0", device_map="auto", torch_dtype=torch.float16)
+
+print("Loading text generation model...")
+# Use a lighter model for testing
+qa_pipeline = pipeline("text-generation", model="gpt2")
 
 # === Qdrant Setup ===
+print("Connecting to Qdrant...")
 qdrant_client = QdrantClient(path="qdrant_data")
 collection_name = "math_problems"
 
@@ -22,16 +29,20 @@ collection_name = "math_problems"
 def is_valid_math_question(text):
     candidate_labels = ["math", "not math"]
     result = classifier(text, candidate_labels)
+    print("Classifier result:", result)
     return result['labels'][0] == "math" and result['scores'][0] > 0.7
 
 # === Retrieval ===
 def retrieve_from_qdrant(query):
+    print("Retrieving context from Qdrant...")
     query_vector = embedding_model.encode(query).tolist()
     hits = qdrant_client.search(collection_name=collection_name, query_vector=query_vector, limit=3)
+    print("Retrieved hits:", hits)
     return [hit.payload for hit in hits] if hits else []
 
 # === Web Search ===
 def web_search_tavily(query):
+    print("Calling Tavily...")
     TAVILY_API_KEY = "your_tavily_api_key"
     response = requests.post(
         "https://api.tavily.com/search",
@@ -48,21 +59,27 @@ class MathAnswer(dspy.Signature):
 # === DSPy Programs ===
 class MathRetrievalQA(dspy.Program):
     def forward(self, question):
+        print("Inside MathRetrievalQA...")
         context_items = retrieve_from_qdrant(question)
         context = "\n".join([item["solution"] for item in context_items if "solution" in item])
+        print("Context for generation:", context)
         if not context:
             return dspy.Output(answer="", retrieved_context="")
         prompt = f"Question: {question}\nContext: {context}\nAnswer:"
-        answer = qa_pipeline(prompt, max_new_tokens=512)[0]["generated_text"]
+        print("Generating answer...")
+        answer = qa_pipeline(prompt, max_new_tokens=100)[0]["generated_text"]
+        print("Generated answer:", answer)
         return dspy.Output(answer=answer, retrieved_context=context)
 
 class WebFallbackQA(dspy.Program):
     def forward(self, question):
+        print("Fallback to Tavily...")
         answer = web_search_tavily(question)
         return dspy.Output(answer=answer, retrieved_context="Tavily")
 
 class MathRouter(dspy.Program):
     def forward(self, question):
+        print("Routing question:", question)
         if not is_valid_math_question(question):
             return dspy.Output(answer="❌ Only math questions are accepted. Please rephrase.", retrieved_context="")
         result = MathRetrievalQA().forward(question)
@@ -79,12 +96,15 @@ def store_feedback(question, answer, feedback, correct_answer):
         "correct_answer": correct_answer,
         "timestamp": str(datetime.now())
     }
+    print("Storing feedback:", entry)
     with open("feedback.json", "a") as f:
         f.write(json.dumps(entry) + "\n")
 
 # === Gradio Functions ===
 def ask_question(question):
+    print("ask_question() called with:", question)
     result = router.forward(question)
+    print("Result:", result)
     return result.answer, question, result.answer
 
 def submit_feedback(question, model_answer, feedback, correct_answer):
@@ -116,4 +136,4 @@ with gr.Blocks() as demo:
                             inputs=[fb_question, fb_answer, fb_like, fb_correct],
                             outputs=[fb_status])
 
-demo.launch()
+demo.launch(share=True, debug=True)