app/graph/nodes/evaluator.py

# # app/graph/nodes/evaluator.py
# from app.core.llm_engine import llm
# from app.core.prompts.evaluator_prompt import evaluator_prompt
# from langchain_core.output_parsers import StrOutputParser
# import json

# chain = evaluator_prompt | llm | StrOutputParser()


# def evaluator_node(state):
#     query = state.get("query")
#     answer = state.get("final_answer")
#     context = state.get("context", "")

#     try:
#         response = chain.invoke({
#                 "query": query,
#                 "answer": answer,
#                 "context": context
#             })

#         # 🔥 clean response (important)
#         response = response.strip()

#         # sometimes model adds ```json
#         if response.startswith("```"):
#             response = response.replace("```json", "").replace("```", "").strip()

#         evaluation = json.loads(response)

#     except Exception as e:
#         print("EVALUATOR ERROR →", e)

#         evaluation = {
#             "relevance_score": 0.5,
#             "context_usage": 0.5,
#             "hallucination": True
#         }

#     return {
#         **state,
#         "evaluation": evaluation
#     }








# app/graph/nodes/evaluator.py
from app.core.llm_engine import llm
from app.core.prompts.evaluator_prompt import evaluator_prompt
from langchain_core.output_parsers import StrOutputParser
import json, re

chain = evaluator_prompt | llm | StrOutputParser()


def evaluator_node(state):
    query = state.get("query")
    answer = state.get("final_answer")
    context = state.get("context", "")

    try:

        response = chain.invoke({
            "query": query,
            "answer": answer,
            "context": context
        }).strip()

        # 🔥 remove markdown/code blocks
        response = re.sub(r"```.*?```", "", response, flags=re.DOTALL).strip()

        # 🔥 extract JSON only
        match = re.search(r"\{.*\}", response, re.DOTALL)

        if match:
            response = match.group(0)

        # 🔥 validate JSON start
        if not response.startswith("{"):
            raise ValueError("Invalid JSON from LLM")

        evaluation = json.loads(response)

        # 🔥 clamp values
        evaluation = {
            "relevance_score": min(max(evaluation.get("relevance_score", 0), 0), 1),
            "context_usage": min(max(evaluation.get("context_usage", 0), 0), 1),
            "hallucination": bool(evaluation.get("hallucination", True))
        }

    except Exception as e:
        print("EVALUATOR ERROR →", e)

        evaluation = {
            "relevance_score": 0.5,
            "context_usage": 0.5,
            "hallucination": True
        }

    return {
        **state,
        "evaluation": evaluation
    }