Skip to content

Mimir v1.3.0 — Phase 3: Offline Embeddings

Choose a tag to compare

View all tags
@tcconnally tcconnally released this 20 Jun 03:50
· 32 commits to main since this release
v1.3.0
256b4c1
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Mimir v1.3.0 — Phase 3: Offline Embeddings

Theme: "Truly zero-dependency semantic search." 10 merged PRs total across 3 phases, 35 MCP tools.

ONNX Local Embeddings

  • bundled-embeddings feature (enabled by default) — ort + tokenizers + ndarray
  • --embedding-model flag wires download of all-MiniLM-L6-v2 from HuggingFace on first use
  • Full fallback chain: LRU cache → local ONNX → Ollama/remote → error
  • Binary stays at 8.8 MB — model is lazy-downloaded, not bundled

Embedding Pipeline

  • In-memory LRU embedding cache (256 entries, Mutex-protected)
  • Incremental re-embedding (skips already-embedded entities)
  • Background embedding available via cron-scheduled mimir_embed calls

SIMD-Accelerated Dense Search

  • ndarray batched dot-product dense search (~10-50x faster at scale)
  • Feature-gated: row-by-row fallback for builds without bundled-embeddings

Install

curl -sSL https://raw.githubusercontent.com/Perseus-Computing-LLC/mimir/main/scripts/bootstrap.sh | bash
Assets 3
Loading