Update app.py

app.py

@@ -19,40 +19,58 @@ model_path = hf_hub_download(
 
 print("🧠 Loading model into llama.cpp engine...")
 # 2. Initialize llama.cpp
-# This is highly optimized C++ code running under the hood
 llm = Llama(
     model_path=model_path,
-    n_ctx=4096,       # Context window (how much conversation it can remember)
+    n_ctx=4096,       # Context window
     n_threads=2,      # Perfectly matches the Hugging Face Free Tier CPU limits
-    verbose=False     # Hides annoying C++ log spam
+    verbose=False     # Hides C++ log spam
 )
 print("✅ Model loaded and ready!")
 
 def generate_answer(message: str) -> str:
+    # 3. Dynamic Prompting Strategy
+    # We use explicit negative constraints to stop database hallucinations natively.
     prompt = (
         "<|user|>\n"
         "You are ChemGist, an intelligent and conversational chemistry assistant. Answer the user's question accurately based on your training.\n\n"
-        "Answer the user's question by weaving the following facts into a natural, flowing paragraph. "
-        "CRITICAL: Do NOT use colons. Do NOT use bullet points. Do NOT output a list. "
-        "Do NOT mention 'Unnamed Compound' or the 'ChEMBL database' in your response. "
-        "Write in complete sentences as if explaining it to a friend.\n\n"
+        "CRITICAL INSTRUCTIONS:\n"
+        "- Focus STRICTLY on the chemical definitions, physical properties, and real-world facts.\n"
+        "- Do NOT mention 'Unnamed Compound', 'ChEMBL', 'database listings', or 'clinical phases'.\n"
+        "- Do NOT use colons, bullet points, or lists.\n"
+        "- Write in complete sentences as if explaining it to a friend, weaving facts into a natural, flowing paragraph.\n\n"
         f"Question: {message}<|end|>\n"
         "<|assistant|>\n"
     )
     
-    # 3. Generate the response using llama.cpp
+    # 4. Generate the response using llama.cpp
     output = llm(
         prompt,
         max_tokens=512,
         temperature=0.3,
         top_p=0.9,
         repeat_penalty=1.1,
-        stop=["<|end|>", "<|user|>"], # Stop generating if it hits these tokens
-        echo=False # Don't output the prompt back to the user
+        stop=["<|end|>", "<|user|>"], 
+        echo=False 
     )
     
-    # Extract the text from the llama.cpp output dictionary
+    # Extract the text
     answer = output['choices'][0]['text'].strip()
+    
+    # 5. Dynamic Sentence Trimming (Post-Processing)
+    # If the response gets cut off and doesn't end with proper punctuation, 
+    # we search backwards to find the last complete sentence and slice the string there.
+    if answer and answer[-1] not in ['.', '!', '?']:
+        last_period = answer.rfind('.')
+        last_exclamation = answer.rfind('!')
+        last_question = answer.rfind('?')
+        
+        # Find whichever punctuation mark came last in the text
+        last_punct = max(last_period, last_exclamation, last_question)
+        
+        # If we found punctuation, cut off the broken fragment after it
+        if last_punct != -1:
+            answer = answer[:last_punct + 1]
+
     return answer
 
 # Build the Gradio Interface