How to fix a stale index: reembed, reenrich, reindex-lexical

The full mechanism and its rationale live in storage.md and storage.md. This guide is only which command fixes which signal.

Three commands force regeneration of derived index state after something outside your notes changed — an embedder or enrichment-LLM upgrade, a cache eviction, or an index-time bug. lode status is what tells you one of them is needed; this page is the depth pointer it links to. If you're deciding whether to run a command, lode status's hint and the command's own --help should already be enough — this page exists for the fuller "what does it actually do" story.

lode reembed

lode status signal: "the index is mixed" (more than one model revision is live) or "the embedder's cached weights have moved past the revision your vectors were embedded with."

What it does: enqueues a fresh embed job for every live head — every note's current version and every external's current snapshot — unconditionally. There's no whole-corpus-vs-targeted flag, because the triggering event (a model or cache change) is itself corpus-wide, not per-note. It rebuilds vectors in place rather than building a parallel index and swapping it in.

Queue vs. execute: queues only. It enqueues the jobs and returns; running them is lode work's job (lode work or lode work --wait). It's resumable by construction — re-running lode work after an interruption picks up where the queue left off, so there's no need to re-run reembed itself.

What it does NOT touch: the enrichment (enrich) leg, and the lexical/FTS leg — the latter is synchronous and model-free, so a model change can't make it stale in the first place. An enrichment-model mismatch is a separate signal handled by reenrich, not this command.

lode reenrich

lode status signal: "the enrichment store's AI annotations disagree with the currently configured enrichment_llm — run lode reenrich."

What it does: enqueues a fresh enrich job, but only for the heads whose stored annotations actually disagree with the currently configured enrichment_llm/provider — targeted, not whole-corpus, unlike reembed. That's a deliberate cost tradeoff: enrichment calls a cloud LLM, so re-running it for the whole corpus on every config change would be needlessly expensive. A head with no annotations at all yet is left to the passive gap-filling reconcile step rather than duplicated here.

Queue vs. execute: queues only, same as reembed — lode work runs the enqueued jobs, and the process is resumable the same way.

What it does NOT touch: the embed queue (never enqueues embed jobs), and it never sweeps in no_egress content — notes/externals marked no_egress are excluded even if their stored annotations happen to be stale, because they're never sent to the cloud enrichment LLM at all.

lode reindex-lexical

lode status signal: "N live note head(s) have no lexical (FTS5) index rows — run lode reindex-lexical to make them keyword-findable now, or wait for lode work to heal them automatically on its next reconcile pass."

What it does: rebuilds the passages_fts lexical index for every live note head from its current body — not just the heads missing rows, so it's a repair tool as much as a backfill. Unlike the embed/enrich staleness signals, a lexical gap self-heals on its own over time via the normal reconcile pass; this command is purely for making it happen now instead of waiting.

Queue vs. execute: synchronous, no queue involved. Chunking and the FTS5 write are the same model-free path the reconcile step already uses in-process; there's no lode work step because there's nothing left to drain once the command returns.

What it does NOT touch: externals. External snapshot FTS rows are written at fetch time by the externals ingest path, not by this command, which only walks notes/versions. Purged notes are the one deliberate exception to "skip purged content": a hard purge already re-indexes the surviving [purged ...] marker body, so this command re-indexes that marker too rather than skipping it — which keeps it faithful to the save path it exists to reproduce. It never re-indexes the purged content itself; that is gone.

Summary

Command	Triggering lode status signal	Scope	Queues (lode work) or executes now
lode reembed	index mixed / revision drift	every live head — notes + externals	queues embed jobs
lode reenrich	enrichment-LLM mismatch	only live heads with stale annotations — notes + externals, excluding no_egress	queues enrich jobs
lode reindex-lexical	missing lexical (FTS5) rows	every live note head — notes only	executes immediately, no queue