Skip to content

feat: add configurable embedding providers#450

Closed
tonyxu-io wants to merge 3 commits into
garrytan:masterfrom
tonyxu-io:feat/copilot-embeddings
Closed

feat: add configurable embedding providers#450
tonyxu-io wants to merge 3 commits into
garrytan:masterfrom
tonyxu-io:feat/copilot-embeddings

Conversation

@tonyxu-io

@tonyxu-io tonyxu-io commented Apr 26, 2026 •
edited
Loading

Copy link
Copy Markdown

Summary

  • Add a generic embedding provider layer with openai, openai-compatible, and copilot providers
  • Configure OpenAI-compatible providers with GBRAIN_EMBEDDING_BASE_URL, GBRAIN_EMBEDDING_API_KEY, GBRAIN_EMBEDDING_MODEL, and GBRAIN_EMBEDDING_DIMENSIONS
  • Keep the Copilot Blackbird shortcut as one provider using the same provider config path
  • Store embedding model metadata with dimensions on import/embed refresh paths
  • Add focused tests for generic provider request formatting, Copilot request formatting, and existing embed behavior

Example

GBRAIN_EMBEDDING_PROVIDER=openai-compatible \
GBRAIN_EMBEDDING_BASE_URL=https://api.example.com/v1 \
GBRAIN_EMBEDDING_API_KEY=... \
GBRAIN_EMBEDDING_MODEL=nomic-embed-text \
GBRAIN_EMBEDDING_DIMENSIONS=768 \
gbrain embed --stale

Test Plan

  • bun run typecheck
  • bun test test/aa-copilot-embedding.test.ts test/embed.test.ts test/import-file.test.ts test/sync-cost-preview.test.ts

@tonyxu-io tonyxu-io force-pushed the feat/copilot-embeddings branch from c09cb59 to 8306aa6 Compare April 26, 2026 14:53
@tonyxu-io tonyxu-io force-pushed the feat/copilot-embeddings branch from 8306aa6 to 263e850 Compare April 26, 2026 14:53
@tonyxu-io tonyxu-io changed the title feat: add Copilot embedding provider feat: add configurable embedding providers Apr 26, 2026
@Cl3MM Cl3MM mentioned this pull request Apr 27, 2026
Closed
@sean-codevasp

sean-codevasp commented Apr 28, 2026

Copy link
Copy Markdown

i was going to try it and you guys already .. haha

@Qodo-Free-For-OSS

Qodo-Free-For-OSS commented Apr 29, 2026

Copy link
Copy Markdown

Hi, The new provider config allows non-1536 embedding dimensions, but the DB schema defines content_chunks.embedding as vector(1536), so inserting 768/1024-length embeddings will fail and vector search will be incompatible.

Severity: action required | Category: correctness

How to fix: Align schema with dimensions

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

Non-1536 embeddings will not fit into content_chunks.embedding vector(1536). This makes Copilot (1024) and any custom dimension (e.g. 768) unusable for persistence/search.

Issue Context

The PR advertises configurable GBRAIN_EMBEDDING_DIMENSIONS and uses it in requests and chunk metadata.

Fix Focus Areas

  • src/core/schema-embedded.ts[92-121]
  • src/core/pglite-schema.ts[71-93]
  • src/core/migrate.ts[41-200] (add migration)
  • src/core/postgres-engine.ts[541-593]
  • src/core/pglite-engine.ts[460-513]

Expected fix

Choose one supported approach:

  1. Keep fixed 1536: disallow configuring other dimensions (validate GBRAIN_EMBEDDING_DIMENSIONS must be 1536; reject Copilot/defaults or add a separate storage path).
  2. Support variable dimensions via migrations: add a migration to ALTER the content_chunks.embedding type to the configured dimension (and rebuild HNSW index), plus ensure PGLite schema/init is consistent.
  3. Store provider-specific embeddings separately: new table/column per dimension/provider.

Also add a guard in upsertChunks (before DB write) to fail fast with a clear error when vector length != expected DB dimension.

Found by Qodo code review

@tonyxu-io

tonyxu-io commented Apr 29, 2026

Copy link
Copy Markdown
Author

Addressed the embedding-dimension storage issue from the review.

Changes:

  • Added migration embedding_dimension_realign to align content_chunks.embedding with configured dimensions, clear stale embeddings when the vector type changes, and rebuild the HNSW index.
  • Passed the configured embedding model/dimensions into schema init before migrations for both Postgres and PGLite.
  • Added upsertChunks fail-fast validation when an embedding vector length does not match the configured dimension.
  • Added regression coverage for 1024-dim Copilot/PGLite schema init, 768-dim config SQL, migration shape, and vector-length guard.

Verified:

  • bun run typecheck
  • bun test test/embedding-dimensions.test.ts test/aa-copilot-embedding.test.ts test/embed.test.ts test/import-file.test.ts test/migrate.test.ts
  • git diff --check

@tonyxu-io
Merge remote-tracking branch 'origin/master' into feat/copilot-embedd…
a5b82d7 
…ings-pr450

# Conflicts:
#	src/core/pglite-engine.ts
#	src/core/postgres-engine.ts
@tonyxu-io

tonyxu-io commented Apr 29, 2026

Copy link
Copy Markdown
Author

Follow-up: rebased/merged the PR branch onto current origin/master and resolved the merge conflicts in the engine init paths.

Current status:

  • PR is now mergeable against master.
  • Kept the upstream forward-reference bootstrap and schema verification paths.
  • Kept the embedding dimension settings/migration path before schema replay, so configurable 1024/768 dimensions still work after the merge.

Verified after the merge update:

  • bun run typecheck
  • bun test test/embedding-dimensions.test.ts test/aa-copilot-embedding.test.ts test/embed.test.ts test/import-file.test.ts test/migrate.test.ts test/schema-bootstrap-coverage.test.ts test/schema-verify.test.ts test/reconcile-links.test.ts test/graph-query.test.ts test/minions-shell.test.ts — 155 pass
  • git diff --check HEAD~1..HEAD

I also tried full bun test; it hit Bun 1.3.9 native segfaults in large-suite runs, while the crash-adjacent files pass standalone (graph-query, minions-shell). No GitHub checks are configured/reported for this branch.

@Cl3MM Cl3MM mentioned this pull request Apr 29, 2026
Closed
@tonyxu-io

tonyxu-io commented May 6, 2026

Copy link
Copy Markdown
Author

Closing this PR because the embedding architecture has changed substantially since this branch was opened.

This branch was based on the old v0.22-era embedding path. Current master uses the v0.27 AI gateway / recipe provider system, so carrying this PR forward as-is would touch the wrong files and create unnecessary conflicts.

I’m opening a fresh, smaller PR against latest master that restores Copilot / Blackbird embeddings using the current provider architecture.

@tonyxu-io tonyxu-io closed this May 6, 2026
@tonyxu-io tonyxu-io mentioned this pull request May 6, 2026
Closed
1 task
tonyxu-io added a commit to tonyxu-io/gbrain that referenced this pull request May 19, 2026
@tonyxu-io
feat: add Copilot embedding provider
5cc285e 
Adds native Copilot embedding provider for GitHub Copilot/Blackbird Metis
embeddings. Calls GitHub /embeddings endpoint directly with the Copilot
request/response shape; not OpenAI-compatible.

- New recipe: src/core/ai/recipes/copilot.ts with metis-1024-I16-Binary
- New implementation kind: 'native-copilot' in types.ts
- gateway.ts: native-copilot branches in instantiateEmbedding/Expansion/Chat
  (embedding does real fetch; chat+expansion throw clear errors since this
  provider is embedding-only)
- embedSubBatch short-circuits when model is a Copilot model, bypassing
  Vercel AI SDK and threading abortSignal into native fetch

Replaces garrytan#450. Built on current master (v0.36) instead of v0.22 legacy path.

Auth: GBRAIN_COPILOT_TOKEN, COPILOT_GITHUB_TOKEN, GH_TOKEN, GITHUB_TOKEN,
or ~/.copilot/config.json (Copilot CLI login).

Config:
  export GBRAIN_EMBEDDING_MODEL=copilot:metis-1024-I16-Binary
  export GBRAIN_EMBEDDING_DIMENSIONS=1024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@tonyxu-io @sean-codevasp @Qodo-Free-For-OSS