EmbeddingGemma-300M as the open-source embedding upgrade

[apocryphx]

What problem does this solve?

First off: nice project! Slicing the YC portfolio by real questions instead of one-company-at-a-time lookup is genuinely useful. And open-sourcing it after the Product Hunt launch is a generous move.

Which brings me to your "Open ideas" list. You mention experimenting with a newer or open-source embedding model for the idea validator. I'd like to suggest a concrete candidate: google/embeddinggemma-300m.

I run it in production as the sole embedder for a local-first semantic memory system (1,300+ document corpus, Core ML on macOS). It replaced BGE-small-en-v1.5 after head-to-head testing.

Here is what I measured:

• English retrieval is at parity with BGE-class models, with quality approaching BGE-M3.
• Genuinely useful multilingual retrieval (100+ languages). In my tests, cross-lingual queries (e.g. Japanese query → English document) worked great. This would quietly serve your internationalization idea too, as YC company descriptions and user queries won't have to stay English-only.
• Zero per-call cost. For your FastAPI backend it's two lines via sentence-transformers. It removes OpenAI from the critical path of the validator and kills the embedding line item entirely.

Two gotchas that are easy to get silently wrong (both cost real accuracy if missed):

1. Asymmetric task prefixes. EmbeddingGemma expects different prompts for queries vs documents (task: search result | query: ... vs title: ... | text: ...). sentence-transformers handles this via prompt_name="query" / "document" — but skip it and retrieval quality drops without any error.
2. Embed the company name/title together with the description. In my benchmarks, including the title in the document embedding was worth +9 points top-1 on cross-lingual retrieval. Company names is exactly where this helps most.

Happy to share more detail on the benchmark setup if useful.

— Kolja (apocryphx)

What do you wish existed?

A free open-source embedding model

References (optional)

Model card: https://huggingface.co/google/embeddinggemma-300m

Scope check

• I checked the Open ideas list and existing issues — this is new.