I am not getting relevant results with m2-bert-80M-32k-retrieval

[legaltextai]

Embedded using the API
Significantly underperforms vs other models
In most of the cases, each embedding is a full text of the Supreme Court decision
Indexed with hnsw.
Should I use a different index?
I store in Postgres and use pgvector for similarity search.
togethercomputer/m2-bert-80M-32k-retrieval
Thanks

[DanFu09]

CC @jonsaadfalcon

Try the V1 models: https://huggingface.co/hazyresearch/M2-BERT-32K-Retrieval-Encoder-V1

Those have seen some legal data during training so hopefully they should work a bit better :)

If they still don't work, would love to hear more about the setup and see if there's a mismatch with how we trained them!

[legaltextai]

both query and the documents use the same embedding protocol, correct? i don't need to add any extra when embedding the prompt, like in UAE Large, right? OK to use cosine similarity and hnsw index? there are some small models, that i am also testing, with smaller context. i doubt they have been trained on any legal data and you can test them and see how they perform.

[DanFu09]

Yes, both queries and documents use the same protocol and model, there's no extra prompt.