Releases · AnswerDotAI/rerankers

Added support for the current state-of-the-art rerankers, BAAI's series of BGE layerwise LLM rerankers, based on Gemma and MiniCPM. These are different from RankGPT, as they're not listwise: the models are repurposed as "cross-encoders", and do output logit scores.

Assets 2

30 Jul 11:34

bclavie

0.4.0

Release bundling up both the 0.3.1 minor release and the 0.4.0 release.

Changes

ColBERT performance improvement! It should now be faster and result in stronger results following implementation of the JaColBERTv2.5 dynamic query length method.
HuggingFace's Text-Embedding-Server (TEI) inference introduced as an API reranker option, thanks to @srisudarsan
T5 bugfix for certain models, where the labels were incorrectly set. Thanks to @marcospiau
Native default support for new Portuguese T5 rerankers introduced by @marcospiau

marcospiau and srisudarsan

Assets 2

17 May 14:17

bclavie

0.3.0

0.3.0 is here! It brings a lot of oft-requested features:

A new transparent Document has been added. You may construct Documents yourself, or keep using the library exactly as-is. This object now allows for metadata support! You can pass a list of metadata (or add them to your Document objects) to rank() calls, and get the metadata back. Thanks to @Anmol6 for starting the work on this!
RankLLM is now supported 🥳 RankZephyr and RankVicuna are implemented, but untested at the moment, while RankLLM + GPT models are fully supported. In version 0.5.0, this will become the default way of using GPT models for reranking purposes.
Some QoL improvements, the most notable of which is that it is now possible to iterate directly on RankedResults objects rather than having to use the wordy for result in results.results.