Updated engine.py to make message payload OpenAI-compatible for Qwen2.5B

engine.py

@@ -128,30 +128,28 @@ import os
 from config import HF_MODEL
 from agents import ValidatorAgent, QuestionGenAgent, ScorerAgent
 
-# Initialize the official serverless client at the module level
+# Initialize the inference client globally using your global configuration string
 client = InferenceClient(model=HF_MODEL, token=os.environ.get("HF_TOKEN"))
 
 def ask_llm(prompt: str, temperature: float = 0.7, max_tokens: int = 512) -> str:
     """
-    Hugging Face Inference Provider API call using the modern chat_completion pipeline.
-    This routes requests via active serverless clusters natively and eliminates task errors.
+    Hugging Face Inference Provider call utilizing the modern chat completion pipeline.
+    This fulfills partner routing rules cleanly and bypasses task metadata limitations.
     """
     try:
-        # Hugging Face API requirements demand a temperature reading strictly > 0
         safe_temp = temperature if temperature > 0 else 0.01
         
-        # ── SWAPPED FROM text_generation TO chat_completion ──
+        # Format the unstructured string prompt into a compliant chat completion message schema
         response = client.chat_completion(
             messages=[{"role": "user", "content": prompt}],
             max_tokens=max_tokens,
             temperature=safe_temp
         )
         
-        # Extract the text string from the standard OpenAI choice payload structure
+        # Extract the content text directly from the returned choices payload array
         cleaned = response.choices[0].message.content
         cleaned = cleaned.replace("<s>", "").replace("</s>", "").strip()
         
-        # Diagnostic print visible inside your Hugging Face Live Logs screen
         print("\n" + "="*40)
         print(f"[LLM RAW OUTPUT]: '{cleaned}'")
         print("="*40 + "\n")
@@ -165,12 +163,11 @@ def ask_llm(prompt: str, temperature: float = 0.7, max_tokens: int = 512) -> str
 
 
 # ── Agent instances (singletons, created once at startup) ─────────────────────
-# These MUST stay directly below ask_llm to prevent initialization NameErrors
+# Retained directly below the function layout to protect against local NameErrors
 _validator   = ValidatorAgent(ask_llm)
 _q_gen       = QuestionGenAgent(ask_llm)
 _scorer      = ScorerAgent(ask_llm)
 
-
 # ── Main orchestration functions (called by Gradio UI) ────────────────────────
 
 def generate_all_questions(job_desc: str, mode_label: str,