Knowledge base

The knowledge base is the “memory” of your AI agents. Upload your price list, service descriptions, FAQ, delivery terms — and the agent will answer customers based on your real facts instead of inventing them. This matters especially for agents in the Inbox: without a knowledge base, AI can sound confident but get prices and details wrong.

The “Knowledge base” section is in the sidebar. You can have several bases — for example, a separate base for each brand or project; switch between them with the selector in the section header.

What you can upload

Source	How to add
PDF	drag the file into the dropzone or pick it from disk
DOCX	same way, the text is extracted automatically
TXT / Markdown	plain text files
Website page	paste the link into the “Add link” field — the page text is pulled in by itself

Files can be uploaded up to 30 MB in size. You can add up to 25 links at once — handy for dropping in all your key website pages at once.

Note

If a page has language versions (for example, the same page in Russian, English, and Uzbek), Repost21 will find them automatically and add them to the base as separate sources. This way the agent can answer customers accurately in different languages.

How it works

1. The uploaded document is split into small chunks — paragraphs and sections.
2. Each chunk is indexed by meaning.
3. When a customer asks a question, the agent searches the base for the most relevant chunks — not by exact word match, but by meaning. The search is multilingual: a question in one language will find facts written in another.
4. The found chunks are passed to the agent as context — and it answers based on them.

You don’t need to configure anything: all processing starts automatically after upload.

File statuses

After being added, a file goes through several stages — the status is shown right in the list:

Status	What it means
in queue	the file is accepted and waiting to be processed
reading	the text is being extracted from the document
processing	the text is being split into chunks and indexed
ready	the file is in the base, the agent can already use it
error	processing failed — try another format or check the file

The whole cycle usually takes less than a minute. You can delete an unneeded file at any time — its chunks will immediately disappear from search.

”Test knowledge” — search test

To the right of the file list there is a “Test knowledge” panel. Ask a question the way a customer would — for example, “how much is delivery?” — and you’ll see which chunks the base finds and how relevant they are.

This is the best way to check the base before launching the agent:

• chunks were found and they answer the question — all good;
• the wrong chunks were found — reword the section in the document more clearly;
• nothing was found — this fact isn’t in the base, add it.

Linking a base to an agent

The base does nothing on its own — you need to link it to an agent:

1. Open the agent in the “AI agents” section and go to the “Knowledge base” tab.
2. Choose an existing base or create a new one right there — the tab has both a dropzone for files and a field for links, so you can fill the base without leaving the agent settings.
3. Save — from this moment the agent searches for answers in the linked base.

Different agents can be linked to different bases: the support agent answers from the FAQ, the content agent relies on product descriptions. More on configuring agents is in the AI agents section.

Strict mode

For agents in the Inbox there is a strict mode: the agent answers only with what was found in the knowledge base. If there’s no answer in the base — it doesn’t make one up but hands the conversation over to a human (escalation) with a polite message to the customer.

Turn on strict mode when the cost of a mistake is high: exact prices, legal terms, medical or financial topics. Without strict mode the agent uses the base as a hint but may supplement the answer with general knowledge.

Tip

Strict mode works together with escalation in the Inbox — the operator will see the conversation the AI couldn’t answer and will continue it themselves.

Tips — how to prepare documents

The cleaner the document, the more accurate the answers:

• One fact — one place. If a price is listed differently in three files, the agent might take any of them. Keep the current version in one document and delete outdated ones.
• Structure the text. Headings, short paragraphs, and lists are split into chunks better than a solid wall of text.
• Phrase it as for a customer. A question-and-answer (“How long does delivery take? — 2–3 days”) is found by search almost perfectly — the FAQ format works best.
• Avoid scans. A text layer is extracted from a PDF; scanned images without text can’t be processed.
• Update the base. Prices changed — upload a fresh file and delete the old one, there’s no need to retrain anything.