LightMem/src/lightmem/memory_toolkits/memory_evaluation.py

from memories.datasets.base import QuestionAnswerPair, MemoryDataset
from inference_utils.operators import (
    QuestionAnsweringOperator,
    LLMExactMatch,
)
from memories import DATASET_MAPPING
import numpy as np
import argparse
import json 
import os 
from typing import (
    List,
    Dict,
    Any,
    Optional,
    Tuple,
)


def _build_context_text(retrieved_memories: List[Dict[str, Any]]) -> str:
    contents = []
    for i, mem in enumerate(retrieved_memories):
        content = mem.get("used_content", '')
        if not isinstance(content, str):
            raise AssertionError("The used_content is not a string for the current memory unit.")
        if not content:
            raise AssertionError("The used_content is empty for the current memory unit.")
        contents.append(f"### Memory {i + 1}:\n{content}")
    return "\n\n".join(contents)


def answer_questions(
    retrievals: List[Dict[str, Any]],
    qa_model: str,
    qa_batch_size: int = 4,
    add_question_timestamp: bool = False,
    dataset_type: Optional[str] = None, 
    interface_kwargs: Optional[Dict[str, Any]] = None,
) -> List[Dict[str, Any]]:
    interface_kwargs = interface_kwargs or {}
    questions: List[str] = []
    contexts: List[str] = []

    for item in retrievals:
        qa_pair: QuestionAnswerPair = item["qa_pair"]
        question_text = qa_pair.question
        if add_question_timestamp:
            question_text = f"{question_text}\nQuestion Timestamp: {qa_pair.get_string_timestamp()}"
        questions.append(question_text)
        base_ctx = _build_context_text(item["retrieved_memories"])
        contexts.append(base_ctx)
    
    
    has_graph = any(
        mem.get("metadata", {}).get("has_graph_relations", False)
        for item in retrievals
        for mem in item["retrieved_memories"]
    )
    if dataset_type is not None:
        dataset_cls = DATASET_MAPPING[dataset_type]
        prompt_name = dataset_cls.get_qa_prompt_name(has_graph) 
    else:
        prompt_name = "question-answering"

    print(f"Using prompt: {prompt_name} for question answering.")
    
    qa_operator = QuestionAnsweringOperator(
        prompt_name=prompt_name,
        model_name=qa_model,
        **interface_kwargs,
    )

    responses = qa_operator(
        questions,
        contexts,
        batch_size=qa_batch_size,
        aggregate=False,
        temperature=0.0,
    )
    return responses

def evaluate_answers(
    retrievals: List[Dict[str, Any]],
    predictions: List[Dict[str, Any]],
    judge_model: str,
    judge_batch_size: int = 4,
    dataset_type: Optional[str] = None, 
    interface_kwargs: Optional[Dict[str, Any]] = None,
) -> List[Dict[str, Any]]:
    interface_kwargs = interface_kwargs or {}
    question_list: List[str] = []
    golden_answers_list: List[List[str]] = []
    prediction_list: List[str] = []
    prompt_name_per_index: List[Tuple[str, str]] = []
    category_per_index: List[Any] = []  
    for i, item in enumerate(retrievals):
        qa_pair: QuestionAnswerPair = item["qa_pair"]
        question_list.append(qa_pair.question)
        golden_answers_list.append([ans for ans in qa_pair.answer_list])
        pred = predictions[i].get("processed_content") 
        # LLMs are not robust to the empty string.
        if pred is None:
            raise ValueError(f"The prediction is None for the question {qa_pair.question}.")
        prediction_list.append(pred)

        category = qa_pair.metadata.get("category", "unknown")
        category_per_index.append(category)

        if dataset_type is not None:
            dataset_cls = DATASET_MAPPING[dataset_type]
            prompt_name, qtype = dataset_cls.get_judge_prompt_info(qa_pair)
        else:
            qtype = qa_pair.metadata.get("question_type", "normal")
            prompt_name = "exact-match"
    
        prompt_name_per_index.append((prompt_name, qtype))

    print(f"Using prompt: {prompt_name} for judgment.")
    judge_operator = LLMExactMatch(
        prompt_name=prompt_name,
        model_name=judge_model,
        **interface_kwargs,
    )

    groups: Dict[Tuple[str, str], List[int]] = {}
    for idx, p in enumerate(prompt_name_per_index):
        if p not in groups:
            groups[p] = [] 
        groups[p].append(idx)

    judge_outputs: List[Optional[Dict[str, Any]]] = [None] * len(retrievals)
    correctness_flags: List[Optional[bool]] = [None] * len(retrievals)

    for (prompt_name, qtype), idx_list in groups.items():
        judge_operator.set_prompt(prompt_name)
        batched_questions = [question_list[i] for i in idx_list]
        batched_golden = [golden_answers_list[i] for i in idx_list]
        batched_predictions = [prediction_list[i] for i in idx_list]
        results = judge_operator(
            batched_questions,
            batched_golden,
            batched_predictions,
            batch_size=judge_batch_size,
            aggregate=False,
            temperature=0.0, 
        )
        for local_pos, global_idx in enumerate(idx_list):
            out = results[local_pos]
            judge_outputs[global_idx] = out
            content = out.get("processed_content")
            if content is None:
                raise ValueError(f"The content is None for the question {batched_questions[local_pos]}.")
            is_correct = "yes" in content.lower() or "correct" in content.lower()
            correctness_flags[global_idx] = is_correct
        # Aggregate each group's results and print the average accuracy
        accuracy = np.mean(
            [correctness_flags[global_idx] for global_idx in idx_list]
        ).item()
        print(f"The accuracy for {qtype} (prompt name: {prompt_name}) is {accuracy:.4f}.")
    
    # Print the overall accuracy
    print("\n" + "="*60)
    accuracy = np.mean(correctness_flags).item()
    print(f"Overall Accuracy: {accuracy:.4f}")
    print("="*60 + "\n")

    finalized = []
    for i in range(len(retrievals)):
        finalized.append(
            {
                "judge_response": judge_outputs[i],
                "is_correct": correctness_flags[i],
                "category": category_per_index[i],  
            }
        )
    return finalized

if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description="A script to evaluate the answers of the search results."
    )
    parser.add_argument(
        "--search-results-path",
        type=str,
        required=True,
        help="Path to the search results."
    )
    parser.add_argument(
        "--qa-model",
        type=str,
        default="gpt-4o-mini",
        help="Model name/path for question answering."
    )
    parser.add_argument(
        "--judge-model",
        type=str,
        default="gpt-4o-mini",
        help="Model name/path for judgment (exact match)."
    )
    parser.add_argument(
        "--qa-batch-size",
        type=int,
        default=4,
        help="Batch size for QA generation."
    )
    parser.add_argument(
        "--judge-batch-size",
        type=int,
        default=4,
        help="Batch size for judge model."
    )
    parser.add_argument(
        "--api-config-path", 
        type=str, 
        default=None,
        help="Path to the API config file."
    )
    parser.add_argument(
        "--dataset-type",
        type=str,
        required=True,
        help="The type of dataset (e.g., LoCoMo, LongMemEval)."
    )
    args = parser.parse_args()

    # Prepare interface kwargs
    interface_kwargs: Dict[str, Any] = {}
    if args.api_config_path is not None:
        with open(args.api_config_path, 'r') as f:
            api_config = json.load(f)
        interface_kwargs["api_keys"] = api_config["api_keys"]
        interface_kwargs["base_urls"] = api_config["base_urls"]
    elif os.environ.get("OPENAI_API_KEY") is not None:
        interface_kwargs["api_keys"] = [os.environ.get("OPENAI_API_KEY")]
        interface_kwargs["base_urls"] = [os.environ.get("OPENAI_API_BASE")]
    
    with open(args.search_results_path, 'r') as f:
        retrievals = json.load(f)
    for item in retrievals:
        item["qa_pair"] = QuestionAnswerPair(**item["qa_pair"])
    print(f"Loaded {len(retrievals)} search results from {args.search_results_path}.")

    # Answer questions
    print("Generating answers with QA model...")
    qa_responses = answer_questions(
        retrievals,
        qa_model=args.qa_model,
        qa_batch_size=args.qa_batch_size,
        dataset_type=args.dataset_type, 
        interface_kwargs=interface_kwargs,
    )

    # Evaluate answers
    print("Evaluating answers with judge model...")
    judge_results = evaluate_answers(
        retrievals,
        qa_responses,
        judge_model=args.judge_model,
        judge_batch_size=args.judge_batch_size,
        dataset_type=args.dataset_type, 
        interface_kwargs=interface_kwargs,
    )

    # Assemble final outputs
    final_results: List[Dict[str, Any]] = []
    for i, item in enumerate(retrievals):
        qa_pair: QuestionAnswerPair = item["qa_pair"]
        ans_dict = qa_responses[i]
        judge_dict = judge_results[i]
        final_results.append(
            {
                "qa_pair": qa_pair.model_dump(),
                "prediction": ans_dict["processed_content"],
                "judge_response": judge_dict["judge_response"],
                "is_correct": judge_dict["is_correct"],
                "retrieved_memories": item["retrieved_memories"],
                "user_id": item["user_id"],
            }
        )

    output_path = args.search_results_path.rsplit('.', 1)[0] + "_evaluation.json"
    with open(
        output_path, 
        'w', 
        encoding="utf-8"
    ) as f:
        json.dump(
            final_results, 
            f, 
            ensure_ascii=False, 
            indent=4, 
        )
    print(f"Saved {len(final_results)} results to {output_path}.")