Knowledge overview

What it is

Knowledge is the material your chatbot is allowed to answer from. Nothing else. It does not browse the open web, it does not know about your company from general training, and it will not invent an answer from thin air — if the fact is not in the knowledge you gave it, the chatbot says it does not know.

That is the single most important idea on this page. Every complaint that starts "the chatbot doesn't know about X" ends at the Knowledge tab.

There are four ways to put material in:

• Text you paste or type directly.
• Files you upload from your device.
• Website crawls, where Agentency fetches pages from a public site.
• Cloud imports from Google Drive or Dropbox, when that is configured.

Whichever route you use, the material becomes a source (also called a dataset): one row on the Knowledge table with a title, a size, and a training status.

How adding knowledge actually works

Understanding the two-step nature of this saves a lot of confused waiting.

Step one: extraction. Agentency pulls the readable text out of whatever you gave it. A .docx gives up its text instantly. A PDF is parsed. A scanned page or a photograph has to be transcribed first, which happens in the background — see Scanned documents and images. A crawled page has its navigation and boilerplate stripped away. Repeated footer contact text (address, hours, phone) is kept once in a Site information source — see Smart content extraction.

Step two: training. The extracted text is broken into passages and indexed so the chatbot can find the right passage for a question. This is what "training" means here. It is not the chatbot learning in a general sense — it is building a searchable index of your material.

Both steps run in the background, after your upload finishes. This is why a source you added five seconds ago is not yet answerable, and why the table shows a status badge rather than just appearing done.

When you would use it

Open Knowledge on day one of a new chatbot, before you share anything with anyone. Come back to it whenever:

• the chatbot answers "I don't know" to something you are sure you documented;
• a price, policy, or opening time has changed;
• you are about to go live and want to confirm the material is trained;
• the usage meter is close to full.

Where to find it

Open Dashboard → Chatbots → your chatbot → Knowledge.

Knowledge belongs to one chatbot. There is no shared library across chatbots — see Collections and ownership for why, and what to do if you need the same material on two chatbots.

Before you start

• The chatbot must already exist. Create it first from Dashboard → Chatbots.
• Your role must allow managing knowledge. A Viewer can read the table but cannot add or remove rows.
• Have your material in a readable form. A password-protected PDF, a page behind a login, and a video are all things Agentency cannot read.

Steps

1. Open Dashboard → Chatbots → your chatbot → Knowledge.
2. Read the usage meter first, so you know how much of your plan's cap is still free.
3. Select Add to open the slide-over drawer. It has a tab per source type: Text, Files, Website, and the cloud cards when configured.
4. Choose the tab that matches your material and fill it in. Every tab asks for a title — make it descriptive, because that title is what the chatbot shows when it cites the source.
5. Leave Start training on unless you are deliberately batching several items to train together.
6. Submit, then stay on the Knowledge table and watch the badge on the new row.
7. When the badge reads trained, open Test Chatbot and ask a question whose answer is only in that source.

What you will see

The Knowledge tab is a table, one row per source, with an empty state when there is nothing yet. Each row shows the title, the collection, the size, and the training badge.

Add opens a slide-over drawer rather than a new page, so you keep your place in the table. Inside the drawer is a live estimate of how much of your knowledge cap the new material will consume.

While training runs, a banner appears above the table. It clears when nothing is actively being processed. A sitemap crawl waiting for its next scheduled scan does not keep the banner spinning — see Crawl from a sitemap.

Choosing the right source type

You have	Use	Why
A short list of facts, FAQs, hours	Text	Fastest, cleanest, no extraction risk
Brochures, manuals, price lists on disk	Files	Handles Office, PDF, spreadsheets
A photo or a scan	Files	Transcribed automatically before training
A public help site or docs tree	Website crawl	No manual copying, repeatable
A big site that publishes a sitemap	Sitemap crawl	Can rescan on a schedule
Documents already in Drive or Dropbox	Cloud import	Skips the download step

When in doubt, start with one small source, confirm the chatbot answers from it, and only then add the rest. Debugging one source is easy; debugging thirty added at once is not.

Limits and plan notes

Knowledge size is capped per chatbot and measured in extracted text, not raw file bytes:

Plan	Knowledge per chatbot
Free	400 KB
Starter	10 MB
Standard	20 MB
Pro	40 MB
Agency	60 MB

Extra knowledge add-ons top this up from a shared pool. A 5 MB PDF full of images might extract to 30 KB of text; a 200 KB plain-text file is 200 KB of text. What counts is the words, not the file. See Knowledge limits.

Training is asynchronous on every plan. There is no "train faster" upgrade — a bigger plan buys you more capacity, not a shorter queue.

Common problems

The table is empty right after I uploaded.

Refresh once, and confirm you are on the same chatbot you uploaded to (the URL contains the chatbot). A brand-new row can take a moment to appear.

Add is blocked with an upgrade prompt.

You are at the knowledge cap for this chatbot. Remove sources you no longer need, or upgrade. See Knowledge limits.

The badge says trained but the chatbot still says it does not know.

Three usual causes, in order of likelihood: you asked using words that never appear in the source; you are testing a different chatbot; or the source extracted almost no text (common with scans and JavaScript-heavy pages). Open the row and check its size — a source showing a few bytes did not really import. See Dataset status and training.

A row has been pending for a long time.

Large files and crawls genuinely take a while. Rows that get stranded are picked up automatically by a background recovery pass, so waiting is usually the right move. If it ends in a failed badge, open the row for the plain-language reason.

I need the same knowledge on two chatbots.

Add it to each one. Sources do not move or attach across chatbots — see Collections and ownership.