The knowledge base lets your AI agents answer with your business's real information (your website, PDFs, Word documents) instead of relying on what they "know" in general. This page explains how it works and why it is done this way.
What it is and how it works
- You give it your content. You enter your website's sitemap (MuseRelay reads only the public pages your
robots.txtallows) and/or upload files: PDF, Word (.docx), text, Markdown or CSV. - MuseRelay splits it up and stores it. Each page or file is divided into fragments of a few sentences. For each fragment it stores the text and a "numeric fingerprint" (a vector) that represents its meaning. Everything lives on our servers in Europe, inside your organisation: no other account can see it.
- When a customer asks, the agent searches. The question is also turned into a fingerprint, and the fragments with the closest meaning (and containing the keywords) are found. Only the 6 most relevant reach the agent, which answers with them and, when helpful, gives the source link.
- It stays up to date. Websites are re-read every day: only what has changed is processed. Uploaded files are not re-read on their own: if they change, delete the old one and upload the new one.
Why a knowledge base and not just the prompt
The "prompt" is the agent's instruction text. It is the right place for short, stable things (who you are, how you speak, what not to do), but not for all of a business's information:
| Prompt only | With a knowledge base | |
|---|---|---|
| How much information fits | A few pages. A huge prompt confuses small models in particular. | Hundreds of pages and documents. |
| Cost per message | All the text travels with every message from every customer, with an AI cost each time. | Only the fragments relevant to that question travel. |
| Updating | By hand: if you change a price or opening hours, you have to remember to edit the prompt. | Automatic: change your website and the agent knows after the next read. |
| Made-up answers | Higher risk: if it is not in the prompt, the model tends to fill the gap. | It answers with what is written in your content, with a source link; if it is not there, it is told to say so. |
| Private documents | They have to be pasted into the prompt. | They are uploaded separately and can be stored in private mode (encrypted). |
In practice they work together: the prompt carries what is stable, the agent's tools bring what changes instantly (for example, free slots in the calendar) and the knowledge base provides your business's "how each thing is done".
Public and private content
- Public (default for websites): the text is stored for search by meaning and by words.
- Private mode (files): the text is stored encrypted and searched by meaning only; it is never shown in the open.
- Do not upload third parties' personal data (customer lists, histories, health data…). The knowledge base is for your business information; you are responsible for the content you upload.
What it costs
- Reading or uploading content uses a small amount of AI credits, only the first time and when a page changes. Re-reading unchanged pages costs nothing.
- Searching the base when answering a customer is not charged separately: what uses credits is the agent's reply, as always.
Limits
- Available from the Basic plan: the free plan does not include it. If your account ends the free Basic trial and moves to the free plan, your bases are kept but your agents stop consulting them and the section is hidden until you return to a paid plan.
- Up to 5 knowledge bases per organisation, each with up to 5 sources and 20 files (max. 10 MB per file).
- Only public pages, normal websites (ports 80 and 443) and those your
robots.txtallows MuseRelay's crawler to read. - The agent only states what is in your content: if something is not written there, it will say it does not know.
If the reading service fails
If the service that computes the fingerprints is unavailable, the chat is never blocked: it keeps answering. Search switches to an emergency mode using exact words in public content (no synonyms or private documents), and the panel shows a notice. New content is saved by words and completed automatically when the service returns.