[FEAT] llm as a judge implementation

vectorq/vectorq_core/similarity_evaluator/strategies/llm_comparison.py

@@ -8,8 +8,35 @@ def __init__(self):
         super().__init__()
 
     def answers_similar(self, a: str, b: str) -> bool:
-        # TODO
-        # @Alex: You can access the inference engine via:
-        # self.inference_engine
-        print("TODO: Embedding-based Answer similarity check not implemented")
-        return False
+        system_message = "You are an expert in determining if two answers are semantically similar."
+        user_prompt = f"""
+Please compare the following two answers:
+<answer_a>
+{a}
+</answer_a>
+<answer_b>
+{b}
+</answer_b>
+Are these two answers semantically similar? 
+IMPORTANT: Respond ONLY with "SIMILAR" or "DIFFERENT".
+"""
+        try:
+            response = self.inference_engine.create(
+                output_format=system_message,
+                prompt=user_prompt,
+            )
+            # Extract the last line for the final decision
+            response_lines = response.strip().split('\n')
+            final_decision = response_lines[-1].strip().upper()
+
+
+            if "SIMILAR" in final_decision:
+                return True
+            elif "DIFFERENT" in final_decision:
+                return False
+            else:
+                print(f"LLM similarity check returned unexpected final decision: '{final_decision}' in response: '{response}'. Defaulting to False.")
+                return False
+        except Exception as e:
+            print(f"Error during LLM similarity check: {e}. Defaulting to False.")
+            return False