VocallieDocs
Open app

Knowledge base

The facts your agent is allowed to state — prices, hours, policies, FAQs. How to load them, and why the agent never invents anything beyond them.

For: Anyone building the agent
On this page

The knowledge base is what your agent knows. During a call, when a caller asks a factual question, the agent searches it, finds the most relevant snippets, and answers from those.

If a fact isn't in here, the agent won't make one up — it'll say it isn't sure and offer to get a human. That's correct behaviour, but it means an empty knowledge base produces an agent that can't help anyone.

Agent → Knowledge

Why it's separate from instructions

You could paste your price list into the agent's instructions. Don't:

  • Everything you put in the instructions is loaded on every call, whether relevant or not. It makes the agent slower and costs more.
  • Long instructions dilute the behavioural guidance the agent actually needs.
  • Knowledge is shared by every agent in your workspace. Write it once, and your sales agent and your support agent both have it.

The knowledge base is searched on demand — only the handful of snippets relevant to the current question reach the agent.

Four ways to load it

  1. 1

    Type a snippet

    A text box and an Add button. Best for one specific fact you just realised is missing.

    Optionally tag it with a collection — a free-text label like pricing, faq, policies — to keep the list organised.

  2. 2

    Upload a CSV

    Bulk import from a spreadsheet. You need a column called document (or text, or content) — that's the snippet the agent will read. An optional collection column tags each row, and every other column is stored alongside as extra detail.

    Rows without one of those three text columns are skipped silently, so check your header row if a count comes back lower than expected.

  3. 3

    Import from your website

    Paste a URL, choose how many pages, and it crawls your site — following links on the same domain only — strips the page down to readable text, breaks it into snippets and loads them.

    This is the fastest way to seed a knowledge base from nothing. It runs while you wait, so keep the page count small (it's capped at 20).

  4. 4

    Delete what's wrong

    Every item has a delete button. Stale facts are worse than missing ones — the agent states them with total confidence.

Careful

Importing the same CSV or URL twice creates duplicates — there's no deduplication. If a crawl went wrong, delete the imported items before re-running it.

What makes a good snippet

The agent retrieves whole snippets, so each one should stand on its own and answer one question completely.

Good:

A standard hygienist appointment is £65 and takes 30 minutes. It includes a scale and polish. New patients should allow an extra 15 minutes for paperwork.

Bad — too fragmented:

Hygienist: £65

Bad — too broad:

Our full price list: check-up £45, hygienist £65, whitening £299, implants from £1,800, orthodontics from £2,400, emergency appointment £80, X-rays included, payment plans available on treatments over £500, deposits required for cosmetic consultations, cancellation within 24 hours forfeits deposit…

That last one gets retrieved for every price question and buries the specific answer. Split it into one snippet per topic.

A checklist

  • One topic per snippet.
  • Include the context, not just the value — "£65" alone is useless out of context.
  • Write it the way you'd say it out loud, since it effectively will be.
  • Use the words your customers use, not your internal terminology. Someone asks about "teeth cleaning", not "prophylaxis".
  • Cover the awkward questions too: parking, accessibility, whether you take their insurance, what happens if they're late.

Start with your actual questions

The highest-value thing you can put in on day one is the answers to the questions you're already asked most. Not what you think is important — what the phone actually asks you.

If you're not sure, run the agent for a week with a thin knowledge base and read the call log. Every "I'm not sure about that" in a transcript is a missing snippet with a customer attached.

How search works, briefly

Two mechanisms, so it degrades gracefully:

  1. Meaning-based search — the primary path. A caller asking "how much to get my teeth cleaned" matches a snippet about hygienist pricing even with no shared words.
  2. Keyword search — the fallback, used if the meaning-based path is unavailable. Less clever, but never leaves the agent stranded mid-call.

The agent is instructed to search before answering any factual question, and it plays a short filler phrase while it does so, so the caller doesn't hear silence.

Limits and gotchas

  • The search box on this page is not the agent's search. It's a plain text filter over the items currently loaded on screen — handy for finding an item to delete, but it doesn't tell you what the agent would retrieve.
  • Occasionally an item can't be processed for meaning-based search (if the AI service is briefly unavailable during import). It's still saved and still findable by keyword, but it won't participate in the smarter search. If a fact seems oddly hard for the agent to find, delete and re-add it.
  • Very long snippets get truncated when processed. Keep each one focused.
  • Collections are free text — a typo creates a new collection. They're for your organisation only; they don't limit what the agent searches.
  • Website imports only see server-rendered content. If your site is a JavaScript app that renders in the browser, the crawler may find little or nothing. Paste the text in manually instead.
  • Your plan sets a knowledge-item ceiling (Trial 50, Starter 1,000, Professional 10,000) shown in your usage.

Keeping it current

Knowledge rots. Prices change, staff leave, a service gets discontinued — and the agent will keep confidently quoting the old version to every caller.

Put a recurring reminder in the diary: once a month, scan the list and check the prices. It takes five minutes and it's the difference between an agent that helps and one that quietly misinforms people.