다중 벡터 검색(ColBERT)

벡터 하나로는 충분하지 않습니다

ColBERT 아이디어

score(q, d) = sum over q_token in q: max over d_token in d: cos(q_token, d_token)

ColBERT가 단일 벡터보다 뛰어난 이유

ColBERT의 비용

ColBERTv2

ColBERT 실행

# pip install ragatouille
from ragatouille import RAGPretrainedModel

rag = RAGPretrainedModel.from_pretrained('colbert-ir/colbertv2.0')
rag.index(collection=documents, index_name='my_corpus')

results = rag.search(query='What is the refund policy?', k=5)

실제 다중 벡터 사용

Hypothetical Questions Pattern

def index_with_hypos(doc):
    # Generate 5 plausible questions this doc could answer
    questions = llm.invoke(f'Write 5 questions answered by this doc:\n{doc}').content.split('\n')
    for q in questions:
        vector_db.add(embed(q), payload={'doc': doc, 'question': q})
# Now queries that match a stored hypothetical question retrieve the doc.

Parent-Child with LangChain

from langchain.retrievers import ParentDocumentRetriever
from langchain.storage import InMemoryStore

store = InMemoryStore()
retriever = ParentDocumentRetriever(
    vectorstore=child_vectorstore,
    docstore=store,
    child_splitter=child_splitter,    # small
    parent_splitter=parent_splitter   # large
)
retriever.add_documents(documents)
# Indexes small chunks, returns large parents.

ColBERT를 사용하는 경우

사용을 건너뛸 경우

BM25와의 하이브리드

ColBERT의 절충점

복습