Implementation:AnswerDotAI RAGatouille RAGPretrainedModel Rerank

Knowledge Sources	RAGatouille RAGatouille Docs
Domains	NLP, Information_Retrieval, Reranking
Last Updated	2026-02-12 12:00 GMT

Overview

Concrete tool for reranking candidate documents using ColBERT late-interaction scoring provided by the RAGatouille library.

Description

The RAGPretrainedModel.rerank() method scores and reorders a set of candidate documents against a query. It delegates to ColBERT.rank() which sets inference max tokens, then calls ColBERT._index_free_retrieve(). This method encodes both query and documents on-the-fly, computes exact MaxSim scores via ColBERT._index_free_search(), and returns sorted results.

The delegation chain:

RAGPretrainedModel.rerank() → delegates to model.rank()
ColBERT.rank() → sets max tokens, delegates to _index_free_retrieve()
ColBERT._index_free_retrieve() → encodes query and docs, calls _index_free_search()
ColBERT._index_free_search() → computes MaxSim scores, returns top-k

Usage

Use to rerank a set of candidate documents from any first-stage retriever. No index required — documents are encoded on-the-fly.

Code Reference

Source Location

Repository: RAGatouille
File: ragatouille/RAGPretrainedModel.py
Lines: L325-357

Signature

def rerank(
    self,
    query: Union[str, list[str]],
    documents: list[str],
    k: int = 10,
    zero_index_ranks: bool = False,
    bsize: Union[Literal["auto"], int] = "auto",
) -> Union[list[dict], list[list[dict]]]:
    """Encode documents and rerank them in-memory.

    Parameters:
        query: Query string or list of queries.
        documents: Documents to rerank.
        k: Number of results per query (default 10).
        zero_index_ranks: Zero-based ranking (default False).
        bsize: Batch size for encoding ("auto" or int).

    Returns:
        list[dict] or list[list[dict]]: Reranked results with keys:
        content, score, rank, result_index.
    """

Import

from ragatouille import RAGPretrainedModel

I/O Contract

Inputs

Name	Type	Required	Description
query	Union[str, list[str]]	Yes	Query string or list of queries
documents	list[str]	Yes	Candidate documents to rerank
k	int	No	Number of top results to return (default 10)
zero_index_ranks	bool	No	Use zero-based ranking (default False, rank 1 = highest)
bsize	Union[Literal["auto"], int]	No	Batch size for encoding (default "auto")

Outputs

Name	Type	Description
return (single query)	list[dict]	Results with keys: content (str), score (float), rank (int), result_index (int)
return (batch queries)	list[list[dict]]	List of result lists, one per query

Usage Examples

Basic Reranking

from ragatouille import RAGPretrainedModel

RAG = RAGPretrainedModel.from_pretrained("colbert-ir/colbertv2.0")

# Candidate documents from a first-stage retriever
candidates = [
    "ColBERT uses contextualized late interaction.",
    "BM25 is a bag-of-words retrieval model.",
    "Late interaction computes token-level similarities.",
    "TF-IDF weights terms by frequency.",
]

results = RAG.rerank(
    query="How does ColBERT work?",
    documents=candidates,
    k=3,
)

for r in results:
    print(f"[{r['rank']}] (score: {r['score']:.4f}) {r['content']}")

Reranking BM25 Results

# After getting BM25 results
bm25_candidates = get_bm25_results(query, k=100)

# Rerank with ColBERT for better precision
reranked = RAG.rerank(
    query="What is late interaction?",
    documents=bm25_candidates,
    k=10,
)

Related Pages

Implements Principle

Principle:AnswerDotAI_RAGatouille_Document_Reranking

Requires Environment

Uses Heuristic

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment