AnswerDotAI / rerankers

A lightweight, low-dependency, unified API to use all common reranking and cross-encoder models.
Apache License 2.0
784 stars 40 forks source link

Render.com fail building wheel for nmslib #14

Open soderalohastrom opened 3 months ago

soderalohastrom commented 3 months ago

Happens when requirements.txt has any of the following: rerankers[gpt] rerankers[rankllm] rerankers[all]

jolackner commented 3 months ago

I have the same issue when my streamlit app requirements.txt has either rerankers[rankllm] or rerankers[all]

bclavie commented 3 months ago

Hey! Thanks guys for flagging. It seems like an old nmslib issue that appeared to have been resolved has returned. I'll try and tweak the requirements to something that works, either a fork of nmslib (as suggested by https://github.com/AnswerDotAI/rerankers/issues/15) or seeing if there's an elegant way to make it install properly again.

bclavie commented 1 month ago

Just as a bit of an update on the current situation with nmslib: it's actually fairly hard to find a workaround for the nmslib install issues on mac, as well as conditionally installing the metabrainz fork -- any workaround seems to result in a slightly different, but just as breaking, issue.

nmslib should be supporting higher versions of python soon, and RankLLM is one of the least used rerankers in this library at the moment, so my plan is probably something as basic as taking it out of [all] requirements and displaying a warning message that it couldn't be installed to nmslib and that you should install the fork if you're on recent python versions, which we'd take away once the new nmslib version is out!