Vidore-benchmark

Changed

- Add the `colpali-engine` dependency
- Add `sentence-transformers` to compulsory dependencies
- Bump ColPali to v1.2 [[model card]](https://huggingface.co/vidore/colpali-v1.2)
- Remove the `baselines` dependency group

Added

- Add int8 quantization for embeddings (experimental)
- Add tests for int8 quantization
- Add dynamic versioning with `hatch-vcs`
- Add loggers in modules (with `loguru`)

Changed

- Re-organize quantization modules

Fixed

- Fix dtype conversion in `ColPaliScorer`
- Fix int8 overflow when computing ColBERT score w/ int8 embeddings
- Fix quantization tests

Removed

- Remove unused `ColPaliWithPoolingRetriever`

Added

- Add experiments on token pooling
- Add `from_multiple_json` method in `EvalManager`

Fixed

- Fix error when using token pooling with bfloat16 tensors
- Add missing L2 normalization in token pooling

Added

- Add support for token pooling in document embeddings

Fixed

- Fix ruff settings

Added

- Add barplot for score per token in the interpretability script

Changed

- [Breaking] All the CLI arguments now clearly show if the batch size is related to the query embedding computation, the doc embedding computation, or the scoring

Fixed

- Fix `DummyRetriever`
- Fix a few broken retriever tests

Fixed

- Fix wrong dictionary key in the `evaluate_retriever` CLI script