Upload documents

An uploaded file becomes a knowledge source: OMAI reads the text out of it, splits it into short passages and stores those passages so the assistant can search them while it answers a customer. Everything happens on Dashboard → Knowledge Base, in the Upload document form.

What you can upload

Five file types are accepted, and the type you pick in File type must match what is actually inside the file. The contents are checked against the type, so a PDF renamed to .docx is rejected rather than half-read.

.pdf, .docx, .txt, .md, .markdown, .csv

The field states the limit for the type you have selected — File (PDF, up to 20MB) and so on — and it changes when you change File type, because there is no single figure: PDF and Word are 20 MB, CSV 10 MB, and TXT and Markdown 5 MB. A file over its own limit is refused in the form with a message naming that limit; nothing is uploaded.

Two things OMAI does not do: there is no way to import a web page or point the assistant at a URL, and text that exists only inside an image is not recognised. A scanned contract, a photographed price list or a PDF exported as images produces no text and fails.

Upload a file

What happens after you upload

The upload itself is immediate; the processing is not. The file is stored and a job is queued, and a background worker picks the job up and runs it through these stages.

1. Reading. The text is extracted from the file. OMAI counts what it managed to read: pages for a PDF, paragraphs for a Word file, rows for a CSV.
2. Checking. The result is classified. A file that produced no usable text fails here. A file where fewer than 90% of the pages or rows were read, or whose table structure could not be preserved, is held for your review instead of being published.
3. Splitting. The text is cut into passages of roughly 500 tokens, never more than 1,200. Headings are repeated into each passage, and a table row, an FAQ entry or a list item is never split down the middle — that is what keeps a price attached to the service it belongs to.
4. Indexing. Each passage is turned into a numeric representation the search can compare against a customer question, in batches of 64.
5. Activation. Only after every passage is confirmed indexed does the new version replace the old one, in a single step. Until that moment the previous version keeps answering.

The page tells you that indexing usually takes a minute or two, and it refreshes itself every 4 seconds while any source is still being processed, so you do not have to reload. There is no fixed processing time, and a very large file takes longer. If a row is still showing Reading document after half an hour, use Re-index on that row.

The status on each source

Every source carries one status badge. Under it sits a line reading, for example, Document · PDF · Version 3 · 42 chunks · updated and the date — the type, the file format, how many times the file has been indexed, and how many passages it currently contributes.

When a document needs your review

A document is held at Needs your review — with the note Review before this goes live — when the reading step produced something OMAI is not willing to publish on its own. The reasons are:

• Fewer than 90% of the pages, paragraphs or rows could be read.
• The character encoding could not be decoded cleanly, so parts of the text may be wrong.
• Table structure could not be preserved reliably, which is how a number ends up next to the wrong label.
• The file contains instruction-like text. It is stored as data only and the assistant will not act on it, but an administrator is asked to look first.

The panel on the row lists exactly what could not be read, and adds a line such as 28 of 34 pages read. Approve and use re-runs the reading step on the same stored file — so if the thing that flagged the document is still in the file, it comes back to Needs your review. The reliable fix is to correct the file itself (export the PDF with a real text layer, flatten the table, remove the odd characters) and upload the corrected copy. Reject archives the source instead, without deleting it.

When a file is refused

Some problems are caught before the file is stored — you get a red line above the form and nothing is added to the list. Others are only visible once the file has been read.

A few of these messages appear in English even when the dashboard is in Hebrew. The wording after The file was rejected: is a fixed code, and the table above tells you what each one means.

Storage allowance

The allowance counts the files you upload. It is a limit on total size, not on how many sources you keep, and the details synced from your settings — services, FAQs, policies and opening hours — have no file behind them and do not use it. Deleting a source releases its share of the allowance as soon as you confirm the deletion; disabling a source does not, because the file is kept.