Upload files

What it is

File upload takes a document from your device, pulls the text out of it, and turns that text into knowledge for one chatbot. The original file is stored, but it is the extracted text that the chatbot searches and quotes.

You can upload several files at once. Each becomes its own source row on the Knowledge table with its own title, size, and training status, so one bad file in a batch does not spoil the rest.

When you would use it

Use uploads when the material already exists as a document: brochures, price lists, product manuals, policy documents, onboarding packs, spreadsheets of specifications. Anything you would email to a colleague who asked the question.

Do not use uploads for material that changes constantly. If your prices live on a web page that you update weekly, a website crawl or a sitemap crawl that rescans on a schedule will stay current on its own; an uploaded PDF will be stale the moment you edit the page.

Where to find it

Open Dashboard → Chatbots → your chatbot → Knowledge → Add → Files.

Before you start

• Know your remaining capacity. The usage meter on the Knowledge tab tells you, and the Add drawer estimates what the new files will cost.
• Check the documents are not password-protected. An encrypted PDF cannot be read and will fail.
• Check they are the right documents. A file uploaded to the wrong chatbot has to be removed and added again — sources do not move between chatbots.

Which file types are accepted

Group	Extensions
Text	.txt
Documents	.pdf, .doc, .docx, .rtf, .odt
Spreadsheets	.xlsx, .xls, .csv, .ods
Images	.png, .jpg, .jpeg, .webp

Anything else is rejected in the picker. The two most commonly missed are PowerPoint (.pptx) and plain HTML files — export a PDF for slides, and use a website crawl for HTML.

Agentency checks the real contents of every file, not just its name. Renaming a .zip to .txt will not sneak it through, and a genuine text file that a scanner mislabels is still accepted.

Images and scanned PDFs take a different path: they have no text to extract, so they are transcribed first. See Scanned documents and images.

Steps

1. Open Dashboard → Chatbots → your chatbot → Knowledge → Add → Files.
2. Enter a title. This appears in the table and in the chatbot's citations, so name it the way a customer would recognise it: "Returns Policy 2026", not "doc3final".
3. Optionally choose a collection to group the source with related material. See Collections and ownership.
4. Drag your files onto the drop zone, or select them from the picker.
5. Check the estimate in the drawer against your remaining capacity.
6. Leave Start training on, then upload.
7. Watch each row on the Knowledge table until its badge reads trained. Scanned documents pass through a Reading Document stage first.
8. Ask Test Chatbot a question using a heading or a phrase taken straight from the document.

What you will see

The Files tab is a drop zone with the same title and collection fields as the other tabs. Files appear in a list as you add them, with a size next to each.

Errors are reported per file. If one file in a batch of six is the wrong type or too large, that file is flagged and the other five still upload. You do not have to start the batch again.

After the upload completes, the drawer closes and you land back on the Knowledge table with the new rows visible and their badges moving.

The upload ceilings

Three separate limits apply to a single upload, and they are not the same as your plan's knowledge cap:

• 10 MB per file. A larger file is rejected before it uploads.
• 20 files per batch. Split a bigger set into several uploads.
• 90 MB per batch in total. The sum of everything in one upload.

These are transfer limits. Your plan's knowledge cap is a separate, text-based limit that applies afterwards — see Knowledge limits. It is entirely normal to be well inside all three upload ceilings and still be blocked by the plan cap, because a 2 MB text-dense PDF can extract to far more text than a 9 MB PDF that is mostly photographs.

Limits and plan notes

• Extracted text — not raw file bytes — counts toward the plan cap: Free 400 KB, Starter 10 MB, Standard 20 MB, Pro 40 MB, Agency 60 MB.
• Very large documents are truncated rather than rejected outright if their text is extreme; in practice you will hit your plan cap long before that.
• Uploading requires a role that permits managing knowledge. A Viewer can see the table only.
• The uploaded file stays stored so you can replace it later, but the chatbot only ever answers from the extracted text.

Getting better answers out of a document

Extraction is literal. It takes what is in the document and nothing more, which means the structure of the document decides the quality of the answers.

• One topic per source, where you can. A single 300-page manual works, but a customer question about returns competes with 299 pages of other content. Splitting the returns chapter into its own source makes it easier to find.
• Headings help. Documents with real headings extract into cleaner passages than one unbroken wall of text.
• Tables are read row by row. A simple price table works well. A complex nested table with merged cells often does not — consider restating the important rows as plain sentences.
• Images inside a document are not read. A price list that exists only as a picture pasted into a Word file will extract as nothing. Type the numbers out.
• Headers, footers, and page numbers come through as noise. Harmless in small doses, but a document where every page repeats a long legal footer wastes a lot of your knowledge cap.

Common problems

The file uploaded but training failed.

Open the row and read the reason — it is written in plain language. The usual causes are a password-protected PDF, a corrupt file, or a scan with nothing readable on it. Retry once; if it fails the same way, the document itself is the problem.

I cannot select a .pptx.

PowerPoint is not accepted. Export the deck as PDF, or paste the speaker notes as text knowledge — often the better option, since slides are mostly bullet fragments that make poor answers.

"File too large."

Over the 10 MB per-file ceiling. Split the document, or export a lighter PDF (most export dialogs have a "reduced size" option that strips embedded images).

The batch failed with a size error but each file is small.

You crossed the 90 MB total for one batch, or the 20-file count. Upload in smaller groups.

The row trained but shows almost no size.

Almost nothing was extracted. This is the signature of a scanned document, an image-only PDF, or a document whose text is all inside pictures. See Scanned documents and images.

I uploaded to the wrong chatbot.

Remove the row there and upload it again on the right chatbot. Sources belong to exactly one chatbot and cannot be moved.

The document is right but the chatbot quotes an old version.

You uploaded the new file as a second source, so both exist. Either remove the old row, or use Replace file on the original row to swap the document while keeping the title and collection.