Skip to content

Releases: AnswerDotAI/byaldi

0.0.5

03 Oct 08:12
Compare
Choose a tag to compare

** Features **

  • Add suport for ColQwen2, the new state-of-the-art model checkpoint in the Col* multi-modal retrieval family, thanks to @jpetrantoni #29

0.0.4

23 Sep 09:31
Compare
Choose a tag to compare

0.0.4

Improvements

  • Backend updated to colpali-engine 0.0.3 in #21, courtesy of Colpali's co-author himself @tonywu71
  • Multi-threaded pdf-to-image conversion with reduced memory usage now supported in #24 thanks to @velaia
  • Langchain integration ported from RAGatouille by @hubkrieb in #23

Fixes

  • Proper support for passing index_root + proper example in the README in #17 by @NathanielLovin
  • Transformers version set to a minimum of 4.42 to minimise compatibility issues
  • Documentation/examples now encourage the use of ColPali-v1.2, which is a strictly superior checkpoint

0.0.3

09 Sep 17:26
8d29b24
Compare
Choose a tag to compare
  • Byaldi can now handle resizing images for you when choosing to store the collection in base64 along with the indexed embeddings, while respecting aspect ratio.
  • Enforcing "fail-first" by returning ValueError rather than easy-to-miss error messages for trying to load an index the wrong way: #8

v0.0.2

06 Sep 11:37
b1d9aa6
Compare
Choose a tag to compare

Minor release addressing a few issues:

  • Indexing whole folder does not adhere to a particular format on every system. To mitigate this, index() and add_to_index() now return a dictionary mapping {doc_id: filename}. Metadata and user-defined doc ids will be supported in a future release.
  • Adding support for non-CUDA accelerators for scoring, thanks to @velaia's work and @ManuelFay and @tonywu71's lighting-fast reactions.
  • Fancy new examples thanks to @comhar

0.0.1

06 Sep 11:24
e789d93
Compare
Choose a tag to compare

World, say hi to Byaldi!