Skip to content

Release Notes: v0.1.6

Date: 2026-05-12

This release bumps the crate version to 0.1.6.

Highlights

  • Bumped crate version from 0.1.5 to 0.1.6.

Post-Release Snapshot

Date: 2026-05-15

This section captures the current main-branch work after the 0.1.6 release. The crate is still on 0.1.6; these notes reflect post-release work that should be visible in the docs and benchmark outputs.

Highlights

  • Added early query routing intent detection for count, sum, and sequence questions.
  • Added routed evidence-first reranking signals for numerical and sequence questions.
  • Added recall_any@k to the scoped LongMemEval benchmark so multi-session reporting matches the benchmark's any-hit interpretation.
  • Added fine-grained query timing fields, including:
  • evidence_ms
  • group_build_ms
  • group_sort_ms
  • lexical_merge_ms
  • posting_scoring_ms
  • routing_seed_ms
  • Optimized routing seeding by removing repeated per-doc text parsing and token-set construction.
  • Reduced query latency substantially on routed benchmark cases after the sparse scoring path was tightened.
  • Hardened persistence paths and the spaCy subprocess script path.
  • Refreshed the README and Tier 1 docs to match the current live query path and the latest benchmark results.
  • Gated POS-derived time hints so they only affect ranking when the query has an explicit temporal phrase.
  • Stopped applying recency bias from verb tense alone to non-temporal queries.
  • Improved the benchmark stability slightly by reducing accidental temporal boosting.

Query Path Notes

  • Query analysis now exposes routing intent for numerical and sequence questions.
  • Routed queries use evidence scoring inside the rerank path.
  • The benchmark and live query paths now both expose richer timing breakdowns for profiling.
  • The current live pipeline remains lexical + sparse + graph based.
  • POS-based time hints remain analysis metadata unless the query is explicitly temporal.

Benchmark Notes

  • Scoped LongMemEval reporting now includes both recall@k and recall_any@k.
  • The current benchmark report used by the README is benchmark/data/lintai_longmemeval_scoped_results.json.