Repository navigation
topic search
Find a discussion even when it uses different words from your question. Search by meaning compares numerical representations of text, called embeddings, produced by a component you select. It needs stored transcripts and an installed local embedding model or an explicitly configured remote service. Use word search immediately if you have not configured one.
A local embedding model processes text on this computer; it must already be installed. A remote endpoint receives transcript text when generating embeddings and your question when searching. Name that endpoint deliberately and choose its model identifier and vector dimensions. The tool does not download models. Previewing generation does not load a model or send transcript text.
The examples below use e5-small for a local model. Confirm it is installed before applying a
batch; its name alone does not install it. Local and remote results need the same model/dimensions
as their stored vectors to be comparable.
Choose an already installed local model, or explicitly select a remote endpoint and dimensions:
zm personal search meetings "what did we decide about pricing?" --model e5-small --since-time 2026-09-01 --json
zm personal search meetings "example plan" --model example-model --base-url http://127.0.0.1:8080/v1 --dims 384 --jsonBefore it reads, the search splits the current transcripts into passages again and embeds every passage that
has no vector for that model yet, so a first search on a large store takes longer and, with a remote model,
sends those passages to it. Continue with --cursor <next_cursor>: the pages continue one ranking.
To prepare vectors ahead of time instead, preview and build them in batches:
zm personal meetings embeddings --model e5-small --json
zm personal meetings embeddings --model e5-small --yes --limit 100 --jsonThe first embedding command previews existing chunk status without opening a model, downloading
files or generating vectors. --yes validates and opens the selected model before rebuilding chunks
and embedding one bounded batch. Local models must already be installed; this command never downloads
them. Remote endpoints are explicit; OpenAI endpoints use OPENAI_API_KEY. Remote models cannot be
used with --offline.
Generation uses current transcript revisions and rechecks their hashes before saving results.
--after-transcript-id continues rebuilding chunks; --after-hash continues missing-vector work.
These are separate cursors. Shared cached vectors remain available to other resources.
A completed generation can emit its acknowledged receipt during a bounded cancellation grace period,
then exit 130. Failed generation and cancelled queries emit no such receipt; a blocked output pipe
can prevent delivery.
The command builds missing stored vectors first, then runs two searches over the same meetings — the date, meeting, series, event, person, project and tag filters apply to both — and ranks them together:
-
By words, as
search meetingswithout--model. A question — a question word first, or a final?— is searched by its content words: "what did we decide about pricing?" by «decide pricing». Other text is the query language; text it refuses is still asked by meaning, andmissingsays so. -
By meaning: every stored passage vector of the model in those meetings, up to
--max-chunks(default 10000).
A line found both ways ranks first and is cited by its own ref; by says how each hit was found, score is the
similarity, and passage gives the first and last line the meaning search matched. On a tie a line that has the
words ranks above a passage that only resembles the question. missing says when --max-chunks stopped the scan.
Index completeness and archive completeness remain unknown. --max-rows and --max-text-bytes bound transcript
reads for generation and search. Over MCP the same command runs with an installed model only: --base-url and
meetings embeddings are not offered to an agent. The measurement
compares words, meaning and both on invented meetings.
Open a ranked passage with meeting context and read its surroundings: similarity is a candidate, not proof of a decision. If there are no candidates, check the selected profile/model, transcript availability and generation/scan coverage before concluding there is no relevant discussion.