README.md

---
language:
- en
license: cc0-1.0
size_categories:
- 10M<n<100m
task_categories:
- text-retrieval
task_ids:
- document-retrieval
---

# abstracts-faiss

This is a [faiss](https://github.com/facebookresearch/faiss) index, trained on [abstracts-embeddings](https://huggingface.co/datasets/colonelwatch/abstracts-embeddings). A ready-to-go search interface for using this index is available at [abstracts-index](https://huggingface.co/spaces/colonelwatch/abstracts-index).

## Building

It was trained with the `train.py` script found at [abstracts-search](https://github.com/colonelwatch/abstracts-search) with the options `-N -c 65536` (normalized, train 65536 clusters), using the default preprocess technique `OPQ96_384` (PCA to a 384-dimensional vector, then apply OPQ for a 96-byte code). Note that, although the Stella model was trained with Matryoshka (MRL) loss, it outputs ordinary vectors which are not expected to be truncated, so PCA was used.

## Tuning

The index comes with the Pareto-optimal parameters from [`faiss.ParameterSpace.explore`](https://faiss.ai/cpp_api/struct/structfaiss_1_1ParameterSpace.html) at `index/params.json`, so a point on the speed-recall tradeoff can be immediately picked. For reference, the `exec_time` field is seconds-per-query on an i7-12700H, and the `recall` field is 1-Recall@1. (1-Recall@1 is the probability that the top result returned by the index is the true closest, measured using a holdout set.) The best recall is 0.756 with a search time of 0.11 seconds, but a recall of 0.723 can be had with a search time of 0.0042 seconds.