Skip to content

MESA v0.9.1 — faster retrieval and article-aware search

Choose a tag to compare

@Yasou13 Yasou13 released this 01 Nov 16:03

MESA v0.9.1 combines Round 13 (speed) and Round 14 (retrieval quality and document workflows), following v0.9.0. SQLite remains the source of truth; LanceDB and Kùzu remain derived projections. Existing V4 HTTP routes and request shapes remain compatible.

Round 13 — speed and operational fixes

  • Cache up to four authorized search scopes, including temporary BM25 indexes and derived text features. SQLite commit changes invalidate the cache; authorization is still checked on every request. MESA_SEARCH_SCOPE_CACHE=false disables reuse.
  • Add ingestion-ledger indexes and reconcile revision activation once per commit transition, keeping existing completeness, supersession and rollback rules.
  • Skip graph seed preparation when graph retrieval is disabled.
  • Move scrypt key verification, issuance and rotation off the event loop without caching verification results.
  • Make mesa add wait for an active revision by default; --no-wait returns after admission. Retry admission capacity errors within the SDK deadline.
  • Accept zero RRF lane weights, preserve timestamped runtime logging through migrations, and record projection start times from outbox claims.

Round 14 — retrieval quality and document workflows

  • Split Turkish legislation at article headings while preserving the exact source text and the 512-token chunk bound. MESA_CHUNKING=auto is the default; plain retains the previous plan for new revisions. Existing active revisions are unchanged.
  • Add optional lazy cross-encoder reranking with BAAI/bge-reranker-v2-m3, after fusion and outside the SQLite search connection. It remains off by default. Install the existing [ml] extra and set MESA_RERANK_ENABLED=true; default depth is 50. Results add rerank_score while retaining fusion scores.
  • Compute query embeddings before taking the shared search connection, allowing provider calls from concurrent searches to overlap.
  • Select default vector RRF weight 8.0 from the frozen Turkish Tier M dev split; other lane weights remain 1.0. Reranker-off passage MRR increased from 0.3953 to 0.5663 under the documented setup. These are corpus-specific dev measurements, not a general performance guarantee.
  • Report terminal chunk failures promptly with mutation ID, state and error class instead of waiting for the document timeout.
  • Delete documents by source with mesa delete --source <ref> or Mesa.delete(source=...); document-ID deletion remains supported.
  • Include OpenAI and Ollama SDKs in the base package. Anthropic remains in [adapters].
  • Add tag-triggered PyPI trusted publishing through GitHub Actions/OIDC.

Install and upgrade

pip install --upgrade https://github.com/Yasou13/MESA/releases/download/v0.9.1/mesa_memory-0.9.1-py3-none-any.whl
mesa start

Requires Python 3.11+. Linux and macOS are supported natively; use Docker or WSL on Windows. Models remain optional and off by default. The command above installs the validated wheel attached to this GitHub release. After PyPI publication is confirmed, the equivalent package-index command is pip install --upgrade mesa-memory==0.9.1. PyPI trusted-publisher setup is a separate owner configuration step.

Follow the backup and upgrade procedure before upgrading an existing store; normal startup migrations add the ingestion indexes. To retain earlier behavior for new revisions, set MESA_CHUNKING=plain; for the previous vector weighting, set MESA_RRF_VECTOR_WEIGHT=1.

An interrupted upload whose admitted prefix matches two different chunking plans fails closed with HTTP 409 rather than silently changing the plan. See configuration.

Evidence and status

  • Both implementation rounds passed local and GitHub checks before merge.
  • Release commit e2f984d4ad97c01782a347e29e44808e5a16dad3: full local gate PASS (1,860 tests and 10 optional MCP tests), 37 focused release tests, metadata preflight, wheel/sdist build and Twine validation PASS. MESA CI and external release gates both PASS.
  • Article coverage: 2,988/3,063 indexed starts (97.55%); all 141 Tier M files round-trip losslessly, with unchanged plain chunks for the 46 non-legislation files.
  • Full measured dev matrix, model/hardware details and limitations: retrieval benchmark.

This is a source release. Production qualification remains NO-GO until the normal certification and production-like gates are complete; CI success and dev retrieval results do not certify production readiness.

Tag-signing status: the owner explicitly approved an unsigned annotated tag for v0.9.1. Metadata preflight passed; signed-tag preflight success is not claimed. The tagged release commit is included in main via PR #141.

Full comparison: v0.9.0...v0.9.1