retriever.py

import pandas as pd
import json
import sys
import os
from collections import defaultdict
from util.vector_base import EmbeddingFunction, get_or_create_vector_base
from util.Embeddings import TextEmb3LargeEmbedding
from langchain_core.documents import Document
from FlagEmbedding import FlagReranker
import time
# from bm25s import BM25, tokenize
import contextlib
import io
from tqdm import tqdm

def rrf(rankings, k = 60):
    res = 0
    for r in rankings:
        res += 1 / (r + k)
    return res

def retriever(requirement, PO, safeguard_vector_store, reranker_model, using_reranker=False, using_BM25=False, using_chroma=True, k=20, if_split_po=True):
    final_result = []
    if not if_split_po:
        final_result = multiretriever(requirement, PO, safeguard_vector_store, reranker_model, using_reranker=using_reranker, using_BM25=using_BM25, using_chroma=using_chroma, k=k)
    else:
        for po in PO:
            po_result = multiretriever(requirement, [po], safeguard_vector_store, reranker_model, using_reranker=using_reranker, using_BM25=using_BM25, using_chroma=using_chroma, k=k)
            for safeguard in po_result:
                final_result.append(safeguard)
    return final_result

def multiretriever(requirement, PO, safeguard_vector_store, reranker_model, using_reranker=True, using_BM25=False, using_chroma=True, k=20):
    """
    requirements_dict: [
        requirement: {
            "PO": [],
            "safeguard": []
        }
    ]
    """
    candidate_safeguards = []
    po_list = [po.lower().rstrip() for po in PO if po]
    if "young users" in po_list and len(po_list) == 1:
        return []
    candidate_safeguards = safeguard_vector_store.get(where={"po": {"$in": po_list}})
    safeguard_dict, safeguard_content = {}, []
    for id, content, metadata in zip(candidate_safeguards['ids'], candidate_safeguards['documents'], candidate_safeguards['metadatas']):
        safeguard_dict[content] = {
            "metadata": metadata,
            "rank": [],
            "rrf_score": 0
        }
        safeguard_content.append(content)
     
    # Reranker
    if using_reranker:
        content_pairs, reranking_rank, reranking_results = [], [], []
        for safeguard in safeguard_content:
            content_pairs.append([requirement, safeguard])
        safeguard_rerank_scores = reranker_model.compute_score(content_pairs)
        for content_pair, score in zip(content_pairs, safeguard_rerank_scores):
            reranking_rank.append((content_pair[1], score))
        reranking_results = sorted(reranking_rank, key=lambda x: x[1], reverse=True)
        for safeguard, score in reranking_results:
            safeguard_dict[safeguard]['rank'].append(reranking_results.index((safeguard, score)) + 1)

    #  BM25
    if using_BM25:
        with contextlib.redirect_stdout(io.StringIO()):
            bm25_retriever = BM25(corpus=safeguard_content)
            bm25_retriever.index(tokenize(safeguard_content))
            bm25_results, scores = bm25_retriever.retrieve(tokenize(requirement), k = len(safeguard_content))
            bm25_retrieval_rank = 1
            for safeguard in bm25_results[0]:
                safeguard_dict[safeguard]['rank'].append(bm25_retrieval_rank)
                bm25_retrieval_rank += 1

        # chroma retrieval
    if using_chroma:
        retrieved_safeguards = safeguard_vector_store.similarity_search_with_score(query=requirement, k=len(candidate_safeguards['ids']), filter={"po": {"$in": po_list}})
        retrieval_rank = 1
        for safeguard in retrieved_safeguards:
            safeguard_dict[safeguard[0].page_content]['rank'].append(retrieval_rank)
            retrieval_rank += 1

    final_result = []
    for safeguard in safeguard_content:
        safeguard_dict[safeguard]['rrf_score'] = rrf(safeguard_dict[safeguard]['rank'])
        final_result.append((safeguard_dict[safeguard]['rrf_score'], safeguard_dict[safeguard]['metadata']['safeguard_number'], safeguard, safeguard_dict[safeguard]['metadata']['po']))
    final_result.sort(key=lambda x: x[0], reverse=True)
        
        # top k 
    topk_final_result = final_result[:k]

    return topk_final_result

if __name__=="__main__":
    embeddingmodel = TextEmb3LargeEmbedding(max_qpm=58)
    embedding = EmbeddingFunction(embeddingmodel)
    safeguard_vector_store = get_or_create_vector_base('safeguard_database', embedding)
    reranker_model = FlagReranker(
        '/root/PTR-LLM/tasks/pcf/model/bge-reranker-v2-m3',
        use_fp16=True,
        devices=["cpu"],   
    )
    requirement = """
    Data Minimization Consent for incompatible purposes: Require consent for additional use of personal information not reasonably necessary to or incompatible with original purpose disclosure.
    """
    PO = ["Data Minimization & Purpose Limitation", "Transparency"]
    final_result = retriever(
            requirement,
            PO,
            safeguard_vector_store,
            reranker_model,
            using_reranker=True,
            using_BM25=False,
            using_chroma=True,
            k=10
            )
    print(final_result)