Feature: [Integration proposal] Octen text and VL embedding functions for LanceDB

SDK

Python

Description

Hello LanceDB team,

We propose registering Octen text and multimodal embedding functions in LanceDB. The text function will preserve query/document semantics; the VL function will make per-item versus fused-vector behavior explicit.

The proposed contribution includes Python support, dimension validation, secret handling through registry variables, configuration reconstruction and a text-to-image retrieval example. Video support would carry explicit media URI and sampling metadata.

We understand that Enterprise embedding generation is client-side. Could maintainers review the proposed registry design and advise whether text and multimodal support should ship in separate PRs?

Background and dated evidence: Octen develops retrieval-focused text and multimodal embedding models. The Octen-Embedding-8B model card reports an RTEB #1 result dated January 12, 2026. Octen's website reports octen-vl-embedding-large at #1 with 80.09 on its May 2026 MMEB-v2 comparison. These are dated results, not a claim of current leaderboard leadership. Hugging Face's API reports 1m+ cumulative downloads across five public Octen model repositories as of September 16, 2026.

References: https://huggingface.co/Octen/Octen-Embedding-8B | https://octen.ai/ | https://huggingface.co/Octen API documentation: https://docs.octen.ai/api-reference/embedding | https://docs.octen.ai/api-reference/vl-embedding Public pricing: https://octen.ai/pricing

Best, Kunka Chief Scientist, Octen [email protected]