8000 Switch embedding model from gemini-embedding-2-preview to stable gemini-embedding-2, and test · Issue #12 · simon-inkie/inkie-the-brain · GitHub
[go: up one dir, main page]

Skip to content

Navigation Menu

Switch embedding model from gemini-embedding-2-preview to stable gemini-embedding-2, and test #12

Open

Description

@simon-inkie

Summary

The brain embeds with gemini-embedding-2-preview (768 dimensions, set in core/config.ts as embeddingModel). A stable gemini-embedding-2 is now listed by the Gemini API. Switch to it and verify.

Why

Preview models can be retired with notice. The stable model avoids that.

Finding so far

Embedding three different sentences with both models at 768 dimensions gave cosine 1.00000 each time, so text vectors look identical and existing vectors should stay valid, with no re-embedding. Only text was tested.

Do

  • Change embeddingModel to gemini-embedding-2.
  • Before merging, embed at least one image, one audio clip and one video clip with both models and compare cosine similarity. If any is below about 0.999, re-embedding the assets collection is needed, so decide before switching.
  • Re-run a handful of known-answer search probes (the 18 Aug weighted-index probes are a good set) and confirm the same top results.
  • Confirm the Qdrant collection dimension (768) is unchanged.

Notes

Agent tooling is under a code freeze until the OSS move, so schedule after that or with an explicit exception.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0