Embedding model used by hindsight cloud

[tempapy]

Hi,
What are the models used in hindsight cloud right now? In particular, what is the embedding model? Hindsight Cloud would be much more practical to me, but my data will be heavily multilingual, and even models ranking high on MTEB can be extremely disappointing in real crosslingual situations. So I'd like to know which models are used so that I can choose between cloud or self hosted (with gemini embedding 2 + probably a gemini flash model, which are the best in my use case).

I didn't find any info on the docs and the slack seems unaccessible, so I'm asking here.
Thank you

[chindris-mihai-alexandru]

The self-hosted default is BAAI/bge-small-en-v1.5, downloaded from HuggingFace on first run. That one matters for your case even if you end up elsewhere, because it is English focused, so heavy multilingual data is exactly where it disappoints.

Embeddings are provider configurable. The models doc lists the supported routes (OpenAI-compatible endpoints, LiteLLM, Vertex, Cohere, Gemini and a few others), and each provider ships its own default model. So self hosting with a Gemini embedding model is a supported path, set the embeddings provider env vars and point them at your key.

What Hindsight Cloud runs internally is not documented anywhere I could find, so nobody here can promise it handles crosslingual recall well. If multilingual recall is the deciding factor, self hosting is the safer call, you keep control of the embedder and can test against your own queries instead of trusting an undocumented default.

On the multilingual point itself, Gemini embedding models are trained multilingual and usually beat the bge family outside English. Worth benchmarking both on a sample of your real data before committing either way.