Skip to content

What it will not do

Eight things Limma does not do, some by design and some not yet. The refusals, error codes and costs an integrator meets are on troubleshooting and errors.

It never guesses

Limma answers only from what your documents and records state, and each fact comes with its source. Where they do not settle a question, the reply says so (possibly, not recorded or undetermined) instead of filling the gap (None versus Not recorded).

It never calls a ranked list complete

A list found by ranking, the passages closest to your words, may hold the right answer, but it is never called complete. Only a counted list says "that's all of them", and it says how many (a count).

There is no substring, wildcard or regular-expression search: a text match has no fixed set to count over, so Limma could never say it found every match. Ask by conditions instead: a field equal to a value (search), or a condition read from each candidate's own documents (criteria).

Audio and video are not read

Recordings are not among the file types Limma reads; a sync skips them and says which files and why (the types it reads).

Scanned pages are read by OCR, in English

A scanned page or an image is read by optical character recognition, in English only.

Connectors only read

A connector reads your source and never writes to it: a bucket needs only list and read permission, Google Drive the read-only scope, SharePoint read access, and a database account is sent SELECT and nothing else (permissions).

It never trains on your data

No model is trained or fine-tuned on your content. What Limma learns from your own use stays in your workspace and is used for nobody else.

Reading takes time

A document is read in the background after it arrives, and a large one takes minutes. Your agent's call waits for the read, 45 seconds by default and at most two minutes; a read that runs longer carries on after the call returns, and asking again says where it is (waiting for the read).