Give your agent a knowledge base

An agent without a knowledge base answers from what its model happens to know, which is fine for how to phrase a sentence and useless for your opening hours. A knowledge base is your own content, searchable by the agent while a call is happening.

Open Knowledge Bases from the sidebar. The page describes itself as domain training sets: files, text, audio, and web pages.

Collections are the unit#

Everything you add lives in a collection, and a collection is what gets attached to an agent. That is the whole organising idea, and it is worth getting right early.

Three counts sit across the top of the page: Collections, Documents and Web sources. They are the fastest way to see whether something you just added actually landed.

The four source types#

New collection, at the top right, opens a menu with four choices. They differ in what they accept, not in how the agent uses them: everything ends up searchable in the same way.

Every dialog asks the same two things#

Whichever type you pick, the dialog opens with a Name at the top and finishes with a Destination collection block at the bottom. Only the middle changes.

The Name is what you will see in the table afterwards. It is worth writing for a future reader rather than accepting whatever the file was called, because a list of documents named export final v3 helps nobody six months later.

Destination collection#

Two radio buttons. Use existing collection is selected by default and gives you a Select Collection dropdown. Create new collection swaps that for a New Collection Name box, so you can start a collection and add its first document in one go.

Adding a file, step by step#

• Click New collection, then File.
• Type a Name.
• Under File Upload, click Choose File. The dialog states the rule itself: only PDF, DOC, and DOCX files are accepted.
• Choose the destination collection, existing or new.
• Click Add file.

Adding text, step by step#

• Click New collection, then Text.
• Type a Name.
• Write or paste into Text Content.
• Choose the destination collection.
• Click Add text.

Adding audio, step by step#

The recording is transcribed first and the transcript is what becomes searchable, so the agent is reading words rather than listening to audio. Transcription happens in the background, which means a long recording will not appear as finished the instant you press the button.

• Click New collection, then Audio.
• Type a Name.
• Under Audio Upload, click Choose File and pick an MP3, WAV, M4A, AAC or OGG.
• Choose the destination collection.
• Click Add audio.

Adding a website, step by step#

The website option has the most settings, so it is worth walking through properly.

• Click New collection, then Website.
• Type a Name. This is what you will recognise it by in the list, not the page title.
• Paste the Website URL and press Check URL next to it. Use it. It tells you now whether the page can be read, rather than after you have saved.
• Decide whether to tick Enable Auto-refresh, which is off by default.
• Choose the destination collection.
• Click Add website.

What auto-refresh does#

Ticking Enable Auto-refresh reveals a Refresh Interval underneath: a number and a unit. It defaults to 3 Days, the number will not go below 1, and the unit can be Days, Weeks, Months or Years. A line underneath confirms your choice back to you, along the lines of URLs being automatically refreshed every 3 days.

This is the setting that keeps a knowledge base honest, and it only exists for websites. A file, some text or an audio transcript stays exactly as it was on the day you added it, forever, until you edit or replace it.

How it actually works#

Worth understanding, because it explains why some content works well and some does not.

Your content is not handed to the agent whole. It is split into chunks, each chunk is turned into a numeric representation of its meaning, and those are stored in a vector index. During a call, the agent turns the caller's question into the same kind of representation and searches for the chunks closest to it in meaning.

That last part is the important bit: the search is by meaning, not by keyword. A caller asking whether you are open on Sunday can match a paragraph headed Trading hours that never uses the word Sunday.

Overlap exists so a sentence that falls on a boundary is not cut in half and lost from both sides. Since you cannot change either number from the dashboard, the way to influence what gets retrieved is to change your content rather than the settings.

What this means for your content#

Because a chunk is what gets found and read, a chunk has to make sense alone. That is the single rule underneath everything below, and it matters more here than in most systems, because you cannot tune your way around it.

• A paragraph that starts with "As mentioned above" is useless once it is retrieved on its own.
• A price in a table with a heading three pages up loses the heading. Put the context in the same chunk.
• One fact in one place. The same price in three documents is three chunks that can disagree.
• Say it plainly. Matching is on meaning, and implied meaning is thin.

Editing a document later#

Each row in the table has a pencil icon. It opens Edit document, showing the document's text in a large editable box, and the dialog is explicit about the consequence: saving re-embeds the new text into the knowledge base.

That is the answer to the obvious worry. A correction is not cosmetic, and you do not need to delete and re-upload to make a fix count.

Attaching a collection to an agent#

Adding content does not by itself change any call. Open the agent, find the Knowledge base card, which describes itself as attaching documents your agent can pull answers from, and select the collection. Until then the agent cannot see any of it, and the card reads None.

This catches people out constantly: they upload a document, test the agent, and it still does not know. Check the attachment before you check anything else.

Finding things once there is a lot#

The table groups rows by collection, with a count next to each collection name and a chevron to fold it away. Each row shows the collection, the type as a coloured badge, the name, when it was last updated and when it was created, with a pencil and a bin under Actions.

Export#

Export downloads a CSV named for the day you took it. It respects whatever filters are active, so filter first and you get exactly the subset you wanted.

One thing to be clear about: it exports the index, not the library. The columns are collection name, collection ID, type, URL, file path, created date, last updated and document ID. Your actual document text is not in the file, so this is a stocktake rather than a backup.