Update PolyAgent/gradio_interface.py

PolyAgent/gradio_interface.py

@@ -1230,24 +1230,44 @@ def llm_only_answer(state: Dict[str, Any], model_name: str, prompt: str) -> str:
 
     client = InferenceClient(model=model_id, token=HF_TOKEN)
 
+    system = (
+        "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."
+    )
+
+    # A simple instruct-style prompt that works for text-generation endpoints
+    flat_prompt = f"{system}\n\nUser:\n{p}\n\nAssistant:\n"
+
     try:
+        # Try chat endpoint first (works only if the provider exposes the model as chat)
         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": "system", "content": system},
                 {"role": "user", "content": p},
             ],
             max_tokens=900,
             temperature=0.7,
         )
         return resp.choices[0].message.content or ""
+
     except Exception as e:
-        return pretty_json({"ok": False, "error": str(e), "model_id": model_id})
+        msg = str(e)
+
+        # If provider says it's not a chat model, fall back to text generation.
+        if ("not a chat model" in msg.lower()) or ("model_not_supported" in msg.lower()):
+            try:
+                out = client.text_generation(
+                    flat_prompt,
+                    max_new_tokens=900,
+                    temperature=0.7,
+                    do_sample=True,
+                    return_full_text=False,
+                )
+                return out if isinstance(out, str) else str(out)
+            except Exception as e2:
+                return pretty_json({"ok": False, "error": str(e2), "model_id": model_id, "mode": "text_generation"})
+
+        return pretty_json({"ok": False, "error": msg, "model_id": model_id, "mode": "chat_completion"})
 
 
 def build_ui() -> gr.Blocks: