Colbert¶
It's higly recommended to use a GPU with ColBERT and to use a fine-tuned model as the base model.
ColBERT can either act as a retriever or as a ranker. While it has the capability to be employed as a retriever, its primary design and strength lie in ranking documents rather than retrieving them. The retriever's key objectives is to eliminate irrelevant documents among a large collection. On the other hand, the ranker is designed to re-order documents (more relevant first) among a small subset of document pre-filtered by a retriever.
In practical terms, if your document collection is relatively small, ColBERT can effectively function as a retriever. However, if you're dealing with a big document collection, ColBERT is better suited to operate as a ranker. In this case, you should use a retriever such as TF IDF, BM25, Sentence Transformers, Splade to pre-filter the documents and then use ColBERT to re-rank the documents.
Colbert ranker with BM25 retriever¶
ColBERT ranker can be used to re-rank candidates in output of a retriever following the code below. We can use a TfIdf retriever, a Splade retriever or a SparseEmbed retriever.
from neural_cherche import models, rank, retrieve
import torch
device = "cuda" if torch.cuda.is_available() else "cpu" # or mps
batch_size = 32
documents = [
{"id": "doc1", "title": "Paris", "text": "Paris is the capital of France."},
{"id": "doc2", "title": "Montreal", "text": "Montreal is the largest city in Quebec."},
{"id": "doc3", "title": "Bordeaux", "text": "Bordeaux in Southwestern France."},
]
retriever = retrieve.BM25(
key="id",
on=["title", "text"],
)
model = models.ColBERT(
model_name_or_path="raphaelsty/neural-cherche-colbert",
device=device,
)
ranker = rank.ColBERT(
key="id",
on=["title", "text"],
model=model
)
retriever_documents_embeddings = retriever.encode_documents(
documents=documents,
)
retriever.add(
documents_embeddings=retriever_documents_embeddings,
)
Once we have created our indexes, we can use the ranker to re-rank the candidates retrieved by the retriever.
queries = [
"What is the capital of France?",
"What is the largest city in Quebec?",
"Where is Bordeaux?",
]
retriever_queries_embeddings = retriever.encode_queries(
queries=queries,
)
ranker_queries_embeddings = ranker.encode_queries(
queries=queries,
batch_size=batch_size,
)
candidates = retriever(
queries_embeddings=retriever_queries_embeddings,
k=1000,
)
# Compute the embeddings of the candidates with the ranker model:
ranker_documents_embeddings = ranker.encode_candidates_documents(
documents=documents,
candidates=candidates,
batch_size=batch_size,
)
scores = ranker(
documents=candidates,
queries_embeddings=ranker_queries_embeddings,
documents_embeddings=ranker_documents_embeddings,
k=100,
batch_size=32,
)
scores
[[{'id': 'doc1', 'similarity': 9.37690258026123},
{'id': 'doc3', 'similarity': 6.458564758300781},
{'id': 'doc2', 'similarity': 6.071964263916016}],
[{'id': 'doc2', 'similarity': 10.65597915649414},
{'id': 'doc3', 'similarity': 6.5705132484436035},
{'id': 'doc1', 'similarity': 5.962393283843994}],
[{'id': 'doc3', 'similarity': 6.877983570098877},
{'id': 'doc2', 'similarity': 4.163510799407959},
{'id': 'doc1', 'similarity': 3.5986523628234863}]]
Note, we could also use the encode_documents method which allows to pre-compute all the embeddings of the documents. This is useful when we have a large number of documents and we want to pre-compute the embeddings of the documents to save time later.