ContextLattice v3.2.13 — Glama-lite

Release date: 2026-03-24

This release upgrades the single-container Glama-lite lane for better retrieval speed and lower tail latency while preserving fail-open behavior and full-stack compatibility.

What Changed

1) SQLite rollup acceleration enabled for lite runtime

• Added a local sqlite topic-rollup index (topic_rollups.sqlite3) built from the live rollup snapshot.
• Enabled sqlite WAL + tuned pragmas (synchronous=NORMAL, temp_store=MEMORY) for faster concurrent read/write behavior.
• Added FTS5 virtual index with BM25 ranking on rollup text fields for faster lexical retrieval in single-container mode.

2) Optional sqlite-vec capability detection lane

• Added runtime detection for fts5 and sqlite-vec module availability via sqlite capabilities.
• sqlite-vec is optional and fail-open: retrieval continues on FTS/in-memory paths if vector extension is not available.
• Added telemetry visibility for lane state under topic rollup health (ftsAvailable/ftsActive, vecAvailable/vecActive).

3) Accuracy guardrails for fail-open correctness

• Topic-rollup search now only uses sqlite index when the sqlite index generation marker matches the in-memory rollup snapshot.
• If markers diverge (for example test/manual in-memory mutation), search falls back to canonical in-memory scoring automatically.

4) Glama-lite profile docs and runtime defaults

• Dockerfile.orchestrator keeps standalone-safe defaults and now explicitly enables sqlite acceleration flags.
• docs/glama-release-compliance.md updated with Glama-lite architecture and relevant env toggles.

Architecture Delta vs Full App

Glama-lite (v3.2.13):

• single-container runtime
• retrieval source default: topic_rollups
• local sqlite outbox + local sqlite rollup acceleration (FTS5 BM25)
• no external service requirements (mongo, mindsdb, pgvector disabled)

Full app (v3 full mode):

• multi-service federated retrieval and fanout
• staged fast lane: topic_rollups + qdrant + postgres_pgvector
• deep async continuation: mindsdb + mongo_raw + letta + memory_bank

Should This Replace Current Lite?

Yes. This should replace the previous lite baseline because it preserves the same standalone contract while materially improving local retrieval behavior and observability without adding external dependencies.

Validation

• python3 -m py_compile services/orchestrator/app.py
• pytest services/orchestrator/tests/test_orchestrator_retrieval.py -k "topic_rollup"

A/B Snapshot (Glama-lite lane)

• Artifact: bench/results/glama_lite_topic_rollup_sqlite_ab_20260324.json
• Dataset: 2,500 topic rollups, 80 searches per lane
• Avg retrieval latency:
  • sqlite FTS lane: 50.225 ms
  • in-memory fallback lane: 85.065 ms
• Observed delta: 1.694x faster (40.96% lower average latency) with sqlite lane active