Update PolyAgent/gradio_interface.py

PolyAgent/gradio_interface.py

@@ -1201,7 +1201,7 @@ def gpt_only_answer(state: Dict[str, Any], prompt: str) -> str:
 # ----------------------------- Other LLMs (Hugging Face Inference) ----------------------------- #
 def llm_only_answer(state: Dict[str, Any], model_name: str, prompt: str) -> str:
     """
-    LLM-only responses using Hugging Face Inference API for non-GPT models.
+    LLM-only responses using Hugging Face Inference API for non-chat models.
     """
     ensure_orch(state)
 
@@ -1242,22 +1242,26 @@ def llm_only_answer(state: Dict[str, Any], model_name: str, prompt: str) -> str:
 
     client = InferenceClient(model=model_id, token=HF_TOKEN)
 
+    # 🔑 Manually format instruction prompt
+    system_prompt = (
+        "You are a polymer R&D assistant. "
+        "Answer directly and clearly. "
+        "Do not call tools or run web searches. "
+        "If you are uncertain, state uncertainty."
+    )
+
+    full_prompt = f"""[INST] {system_prompt}
+{p} [/INST]"""
+
     try:
-        resp = client.chat_completion(
-            messages=[
-                {
-                    "role": "system",
-                    "content": (
-                        "You are a polymer R&D assistant. Answer directly and clearly. "
-                        "Do not call tools or run web searches. If you are uncertain, state uncertainty."
-                    ),
-                },
-                {"role": "user", "content": p},
-            ],
-            max_tokens=900,
+        resp = client.text_generation(
+            full_prompt,
+            max_new_tokens=900,
             temperature=0.7,
+            do_sample=True,
+            return_full_text=False,
         )
-        return resp.choices[0].message.content or ""
+        return resp.strip()
     except Exception as e:
         return pretty_json({"ok": False, "error": str(e), "model_id": model_id})