コンテンツにスキップ

Guide

Cookbook

Reference

Concepts

reyn embeddings

Inspect and manage the action embedding index that backs search_actions. Reads on-disk state only — no network probes; matches the cheap-by-default posture of reyn mcp list.

The action index rides the same pluggable IndexBackend doc-RAG sources use (one cosine implementation, one advisory-lock, one storage/dedup path — unified in FP-0057 Phase 0); its cache lives at .reyn/cache/index/actions/ alongside other indexed sources. Pre-consolidation it had a separate, hand-rolled SQLite schema and cosine implementation under .reyn/cache/action_index/ — that path is no longer read or written (clean-break; the cache is fully regenerable, so no migration was needed).

See Concepts: enable semantic search for the onboarding walkthrough, and Concepts: RAG — Embedding configuration for the underlying embedding.classes map.

Synopsis

reyn embeddings status [--json]
reyn embeddings rebuild [<class_name>]
reyn embeddings clear

Subcommands

status [--json]

Show one row per configured embedding class with name / backend / model / cache_path / size_mb / indexed_actions / last_built. Reads reyn.yaml for the class list and .reyn/cache/index/actions/index.db for the current on-disk binding.

$ reyn embeddings status

NAME      BACKEND  MODEL                           CACHE_PATH                  SIZE_MB  ACTIONS  LAST_BUILT
────────────────────────────────────────────────────────────────────────────────────────────────────────────
light     litellm  openai/text-embedding-3-small   .reyn/cache/index/actions      0.31       87  (never)
standard  litellm  openai/text-embedding-3-small   .reyn/cache/index/actions      0.31       87  2026-05-27T19:02:00+00:00
strong    litellm  openai/text-embedding-3-large   .reyn/cache/index/actions      0.31        0  (never)

Every class's backend reads litellm (#3128 removed reyn's in-process sentence-transformers backend — reyn depends on litellm exclusively for embeddings now, whether the model runs at a provider's own API or behind an operator-run litellm proxy). cache_path / size_mb are the shared .reyn/cache/index/actions/ SQLite — not a per-class download cache; there is nothing left for this command to manage other than that index. The SQLite cache stores one model_class at a time (= one row carries the current indexed_actions / last_built; the rest report 0 / "(never)" for the build state, though size_mb is the same shared directory for every row). This avoids ambiguity about which class owns the on-disk index after a class swap.

--json emits the same data as a list of dicts for scripting:

reyn embeddings status --json | jq '.[] | select(.indexed_actions > 0)'

rebuild [<class_name>]

Drop the on-disk action index SQLite + WAL sidecars + .build.lock marker so the next reyn chat session re-embeds. Does NOT itself trigger embedding; the next Session that uses search_actions re-runs ActionEmbeddingIndex.build().

reyn embeddings rebuild
# removed action index cache: index.db, index.db-wal.
# The next chat session will re-embed.

A clean project state (= no .reyn/cache/index/actions/index.db) reports "nothing to rebuild" without erroring:

reyn embeddings rebuild
# no action index found at .reyn/cache/index/actions/index.db; nothing to rebuild.

<class_name> is accepted for forward compatibility with a future per-class cache. Today the SQLite layout stores one model_class at a time, so passing a name verifies the class exists in reyn.yaml and notes the cache-shared semantics:

reyn embeddings rebuild standard
# removed action index cache (shared across classes today): index.db, index.db-wal.
# The next chat session will re-embed for class 'standard'.

Unknown class names exit non-zero with the configured list inline:

reyn embeddings rebuild stnadard   # typo
# error: embedding class 'stnadard' not in reyn.yaml's embedding.classes.
# Configured: ['light', 'standard', 'strong']

clear

Removes .reyn/cache/index/actions/ (= SQLite index + WAL sidecars + build lock). #3128 removed reyn's in-process sentence-transformers backend, so clear no longer also wipes a downloaded-model cache — the SQLite action index is the only on-disk state this command manages; every embedding call routes through litellm (a provider's own API, or an operator-run litellm proxy), which owns any model cache on its own side, outside reyn's .reyn/cache/ tree entirely.

reyn embeddings clear
# removed /path/to/.reyn/cache/index/actions
# freed ~0.31 MB

Useful for:

  • Cache corruption — a partial write / interrupted build leaves the SQLite cache in a state that the next embed() call can't recover from on its own.
  • Reclaiming disk / starting fresh — the index is fully regenerable (content-hash re-embeds only what changed), so wiping it is always safe; the next search_actions build re-creates it.

Absent paths are reported as skipped — calling clear on a clean install is a no-op.

On this page