Skip to content

v1.23.0: pluggable embedding providers (bring a frontier embedder)

Latest
Latest

Choose a tag to compare

View all tags
@kitfunso kitfunso released this 08 Jun 19:30
· 14 commits to master since this release
v1.23.0
fb96f7e
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Added

  • Pluggable embedding providers. config.embeddings.provider now accepts local (default, zero-dependency @xenova/transformers) or an opt-in API provider: openai, voyage, or cohere. Bring a frontier embedder such as OpenAI text-embedding-3-large for frontier-class retrieval, while the default stays fully local with no API keys and no new runtime dependencies (API providers use the native fetch on Node 22.5+). The key is read from OPENAI_API_KEY / VOYAGE_API_KEY / COHERE_API_KEY. Optional embeddings.apiBaseUrl (https only; localhost may use http) and embeddings.batchSize (positive integer, default 64). API requests are bounded by a 30s timeout.

Changed

  • Provider-aware reindex identity. The local provider keeps the bare model string as its identity, so existing stores are NOT re-indexed on upgrade. Switching to or from an API embedder (or changing the model or vector dimension) triggers the existing reindex path.
  • embeddings.enabled: false now hard-disables embedding for both local and API providers, which prevents unwanted paid API calls.
  • Status surface. hippo status and hippo embed --status now show the active provider, model, whether the API key is present, the vector dimension, and a reindex-pending hint, and still report cached counts when embeddings are disabled or a key was removed.

Notes

  • Embedding stays best-effort on the add path: a provider failure never rejects a write. The explicit hippo embed backfill checkpoints progress per chunk and surfaces hard failures, so it never reports a false success. A malformed API response (wrong vector count or an empty vector) is treated as a hard failure so a reindex aborts atomically.
  • A real frontier-embedder LongMemEval benchmark (dual numbers: the zero-dependency local floor plus the frontier number) is a follow-up that must run on a host with API egress and a key. See docs/FRONTIER_EMBEDDER_BENCHMARK.md.

Reviewed across a plan-stage senior-code-review and 10 rounds of cross-model codex review, all findings fixed. Full test suite green (2525 tests).

Assets 2
Loading