fast-forward-indexes 0.1.0

Efficient interpolation-based ranking on CPUs

Fast-Forward Indexes

This is the reference implementation of Fast-Forward indexes.

⚠ Important: As this library is still in its early stages, the API is subject to change! ⚠

Installation

Install the package via pip:

pip install fast-forward-indexes

Getting Started

Using a Fast-Forward index is as simple as providing a TREC run with sparse scores:

from pathlib import Path
from fast_forward.encoder import TCTColBERTQueryEncoder
from fast_forward.index import InMemoryIndex, Mode
from fast_forward.ranking import Ranking

# choose a pre-trained query encoder
encoder = TCTColBERTQueryEncoder("castorini/tct_colbert-msmarco")

# load an index from disk into memory
index = InMemoryIndex.from_disk(Path("/path/to/index"), encoder, Mode.MAXP)

# load a sparse run (TREC format)
sparse_ranking = Ranking.from_file(Path("/path/to/sparse/run.tsv"))

# load all required queries
queries = {
    "q1": "query 1",
    "q2": "query 2",
    # ...
    "qn": "query n"
}

# compute the corresponding dense scores and interpolate
alpha = 0.2
result = index.get_scores(
    sparse_ranking,
    queries,
    alpha=alpha,
    cutoff=10,
    early_stopping=True
)

# create a new TREC runfile with the interpolated ranking
result[alpha].save(Path("/path/to/interpolated/run.tsv"))

Documentation

A more detailed documentation is available here.

Examples

• Creating a Fast-Forward index from a prebuilt Pyserini index
• Computing dense scores for a sparse run and interpolating