Update model ID and refine SQL generation process in app.py for improved accuracy and clarity

app.py

@@ -4,8 +4,8 @@ from transformers import T5Tokenizer, T5ForConditionalGeneration
 from sentence_transformers import SentenceTransformer, util
 
 # --- CONFIGURATION ---
-# Ensure this matches your ACTUAL model on Hugging Face
-FINE_TUNED_MODEL_ID = "hmyunis/t5-sql-finetuned"
+# UPDATE THIS to the new model you just trained
+FINE_TUNED_MODEL_ID = "hmyunis/t5-base-sql-custom" 
 
 print(f"Loading Model: {FINE_TUNED_MODEL_ID}...")
 try:
@@ -17,46 +17,43 @@ except Exception as e:
     print(f"CRITICAL ERROR LOADING MODELS: {e}")
 
 def get_sql_pipeline(question, all_columns_str):
-    print(f"Received Question: {question}")
-    print(f"Received Columns Length: {len(all_columns_str)}")
-
+    print(f"Input Q: {question}")
+    
     try:
         # 1. Parse Columns
-        all_columns = eval(all_columns_str)
-
+        all_columns = eval(all_columns_str) 
+        
         # 2. Schema Linking (Embeddings)
         question_embedding = embedder.encode(question, convert_to_tensor=True)
         column_embeddings = embedder.encode(all_columns, convert_to_tensor=True)
         hits = util.semantic_search(question_embedding, column_embeddings, top_k=6)
         relevant_cols = [all_columns[hit['corpus_id']] for hit in hits[0]]
-
-        # 3. Formulate Prompt
+        
+        # 3. Formulate Prompt (MATCHES TRAINING EXACTLY)
         schema_context = ", ".join(relevant_cols)
-        input_text = f"translate to SQL: {question} </s> {schema_context}"
-        print(f"Prompt sent to T5: {input_text}")
-
+        # Note the prefix change: "translate English to SQL"
+        input_text = f"translate English to SQL: {question} </s> {schema_context}"
+        print(f"Prompt: {input_text}")
+        
         # 4. Generate
         input_ids = tokenizer(input_text, return_tensors="pt").input_ids
-        outputs = model.generate(input_ids, max_length=128)
+        
+        # Use beam search for better accuracy (slower but worth it)
+        outputs = model.generate(
+            input_ids, 
+            max_length=128, 
+            num_beams=4, # Inspects 4 possible paths
+            early_stopping=True
+        )
+        
         generated_sql = tokenizer.decode(outputs[0], skip_special_tokens=True)
-
-        print(f"Generated Output: '{generated_sql}'")
-
-        # Fallback if empty (Model produced nothing)
-        if not generated_sql:
-            return "SELECT * FROM api_customer -- Model returned empty, defaulting."
-
+        print(f"Output: '{generated_sql}'")
+        
         return generated_sql
 
     except Exception as e:
-        error_msg = f"Error in HF Space: {str(e)}"
-        print(error_msg)
-        return error_msg
+        return f"Error: {str(e)}"
 
-# Launch Gradio
-iface = gr.Interface(
-    fn=get_sql_pipeline,
-    inputs=["text", "text"],
-    outputs="text"
-)
-iface.launch()
+# Simpler Interface
+iface = gr.Interface(fn=get_sql_pipeline, inputs=["text", "text"], outputs="text")
+iface.launch()