Skip to content

Recall 2.6.0 — better hybrid retrieval, unchanged lexical default

Latest

Choose a tag to compare

@bledden bledden released this 09 Sep 03:53
· 3 commits to main since this release

Recall 2.6 improves optional hybrid retrieval while preserving the default lexical behavior from 2.5.

  • Strong lexical or semantic matches survive rank fusion instead of being buried by weaker overlapping results.
  • Results retain the passage that earned its rank. Overlapping matches from the same retained revision are combined into an exact source span, capped at 3,200 characters.
  • No new required model, dependency, network request or schema migration. Embeddings remain opt-in, and no configured model is changed automatically.

On a frozen 25-positive real-history evaluation, BGE hybrid exact hit@5 improved from 20→23 unscoped and 22→24 scoped; Arctic XS improved from 19→21 and 20→23. Neither model lost a previously passing exact or broad-evidence case. Default lexical results remained 24/25 in both scopes. These are limited local results, not general competitor superiority or token-saving claims. A development-only question-ranking hint was rejected after real-history regressions and is not included.

Validation: 908 source tests on Python 3.9 and 3.12, independent review on Python 3.9 and 3.14, plus real-history per-case gates and packaged-reader checks.

Updating

Update your Recall plugin/runtime using your existing installation route. Store schema stays at 12; existing history and vectors need no rebuild. Lexical-only Claude Desktop/Cowork readers are unaffected by this fix. Running Claude Code sessions can finish with their existing cache; use a new session or the host's reload workflow to load the new plugin version.

Download the curated ZIP and verify it with the attached SHA-256 file. Community-catalog pins may lag the release; no resubmission is needed.

Installation and updates · Changelog

Co-authorship and review

Co-authored with Codex, contributing the hybrid retrieval and evidence-preservation implementation, regression/evaluation tooling, documentation, and release preparation. Blake Ledden directed the work and approved publication; Fable provided independent review.