Knowledge Base
On this page
A Knowledge Base is the binder on the desk. The agent does not memorise it — it looks things up in it, mid-call, when a caller asks something it needs.
When you need one (and when you do not)
The dividing line is simple, and getting it right keeps your agent both fast and accurate.
| Put it in… | Because… | Examples |
|---|---|---|
| Instructions | It applies to every call and shapes behaviour. Re-read on every reply. | Tone, escalation rules, opening line, the five prices you are asked about constantly |
| Knowledge Base | It might be asked about. Searched only when relevant, so size costs nothing. | A 400-line price list, refund policy, product manual, coverage tables, the staff handbook |
If you find yourself pasting a document into the instructions box, stop — that is the signal. It will slow every call down and the agent follows long prompts less reliably.
Creating one
Open Knowledge Base and create
Go to Knowledge BaseNew and give it a name. Name it after its contents —
Pricing 2026,Policies & refunds— not after the agent, because one knowledge base can serve several agents.Add your sources
Upload files, paste web page addresses, or both. Each source is processed in the background; large PDFs take a minute or two. You will see file and page counts on the knowledge base once it is ready.
Attach it to the agents that need it
A knowledge base does nothing until an agent has it. Attach from the agent, then tell the agent in its instructions that it exists — see below.
Test with a question only the document can answer
Pick something obscure from page 30. If the agent answers it correctly, the pipeline works. If it says it does not know, the source probably has not finished processing — or it is a scan, which is covered below.
What you can add
Files
.pdf, .docx, .txt and .csv, up to 10 MB each. Your plan sets how many knowledge bases you can have and the total storage across them.
If a PDF is a photograph of a page, there is no text in it to read — the agent will find nothing. Test by opening the file and trying to select a sentence with your cursor. If you cannot, it is a scan. Re-export it from the original document, or retype the important parts into a .txt file.
Web pages
Paste a URL and Vociply fetches the page, strips the navigation and adverts, and keeps the content. Point it at your pricing page, FAQ or service pages.
A page is captured as it was on the day you added it. If you change the page, re-add it — it does not re-fetch on its own.
Writing documents an agent can actually use
Documents written for humans to skim are often poor for retrieval, because the agent pulls out passages rather than reading front to back. A few habits make a large difference.
- Make every section self-contained. A heading that says “Standard rate: KSh 3,500” beats a bare table cell under a heading three pages earlier, because the retrieved passage carries its own context.
- Write the question, then the answer. FAQ format retrieves extremely well — the caller’s phrasing tends to match your heading.
- Spell out what you do not do. “We do not offer orthodontics” is a fact worth storing. Without it the agent has nothing to retrieve and is more likely to improvise.
- Avoid “see above” and “as mentioned”. There is no above when a single passage is pulled out on its own.
- Keep one topic per document. Pricing, policies and product specs as three documents beats one 90-page everything file.
- Put the date in the document. “Prices effective March 2026” lets the agent say how current the figure is.
Wide pricing grids often lose their meaning once extracted from a PDF — row and column headers get separated from the numbers. If a table matters, also write it out as sentences: “A 3-bedroom deep clean in Nairobi costs KSh 8,000 and takes about four hours.” Ugly, and it works.
Attaching it to an agent
Two steps, and the second is the one people forget.
Attach the knowledge base
In AgentsEdit, select the knowledge base. One agent can have more than one, and one knowledge base can serve many agents.
Tell the agent it exists
Add a line to the instructions along these lines:
You have access to our price list and policy documents. Search them before answering any question about prices, coverage or policy. Never state a price that is not in them.Without this, an agent that has a knowledge base may still answer from its own assumptions. The instruction is what makes it reach for the binder.
Keeping it accurate
An out-of-date knowledge base is worse than none, because the agent will state old prices with total confidence.
- Re-upload the document whenever the underlying facts change. Remove the old one.
- Re-add web pages after you edit them — the captured copy does not update by itself.
- When something changes, test it: ask the agent the question directly and confirm you hear the new answer.
- Diarise a quarterly review of anything with a price or a date in it. Fifteen minutes.
Knowledge Base questions
How many knowledge bases can I have?
It depends on your plan — both the number of knowledge bases and total storage are capped. Billing shows your current limits, and the app tells you when you are at one.
Does a big knowledge base slow down calls?
No, not in the way you would expect. Only the passages relevant to the question are retrieved, so a 300-page manual costs nothing on a call that never asks about it. Long instructions slow down every call; a long knowledge base does not.
Can the agent read out a whole document?
No, and you would not want it to. It retrieves the relevant passage and answers in a sentence or two. If a caller genuinely needs the whole document, have the agent offer to email or text it — see Tools.
It is not finding something I know is in there.
In order of likelihood: the source is still processing; the PDF is a scan with no text layer; the passage relies on context from elsewhere in the document (“see above”); or the instructions never told the agent to search. Work through those four and it is almost always one of them.
Can I put customer data in a knowledge base?
You can, but you usually should not. A knowledge base is shared reference material for every caller who reaches that agent. For per-caller data — this customer’s order, this member’s balance — use an integration that looks the person up live.