교차 인코더를 사용한 재순위화

재순위화하는 이유

이중 인코더와 교차 인코더

2단계 검색

Cohere Rerank 사용

import cohere
co = cohere.Client(COHERE_KEY)

results = co.rerank(
    model='rerank-multilingual-v3.0',
    query='What is the refund policy?',
    documents=top_50_chunks,
    top_n=5
)
for r in results.results:
    print(r.index, r.relevance_score, top_50_chunks[r.index])

오픈 소스 재순위화기

from sentence_transformers import CrossEncoder
model = CrossEncoder('BAAI/bge-reranker-base')

pairs = [(query, chunk) for chunk in candidates]
scores = model.predict(pairs)
ranked = [c for _, c in sorted(zip(scores, candidates), reverse=True)][:5]

Voyage Rerank

import voyageai
vo = voyageai.Client(api_key=VOYAGE_KEY)
res = vo.rerank(
    query=query,
    documents=candidates,
    model='rerank-2',
    top_k=5
)

재순위화가 도움이 되는 경우

재순위화가 필요하지 않은 경우

비용 절충

Integrating into LangChain

from langchain.retrievers import ContextualCompressionRetriever
from langchain.retrievers.document_compressors import CohereRerank

compressor = CohereRerank(top_n=5)
retriever = ContextualCompressionRetriever(
    base_compressor=compressor,
    base_retriever=vector_retriever
)

Integrating into LlamaIndex

from llama_index.postprocessor.cohere_rerank import CohereRerank

rerank = CohereRerank(top_n=5)
query_engine = index.as_query_engine(
    similarity_top_k=20,
    node_postprocessors=[rerank]
)

다양성을 고려한 재순위화

from langchain.vectorstores import Chroma
results = store.max_marginal_relevance_search(query, k=5, fetch_k=20, lambda_mult=0.5)

2단계 패턴

복습