Knowledge base

Ground your agents in your own content so they answer from your documents instead of guessing.

A knowledge base is a folder of your content - your FAQs, your policies, your product details - that an agent can look things up in while it’s on a call. Attach one to an agent and, the moment a caller asks something specific (“what’s your refund window?”, “do you deliver to Pune?”), the agent finds the relevant passage in your material and answers from that - instead of improvising a confident-sounding guess.

That last part is the whole point. A language model without a knowledge base will happily invent an answer that sounds right. On a real call, with a real customer, a wrong-but-confident answer is worse than “let me check.” A knowledge base is how you stop the guessing: the agent answers from what you actually wrote, and when your documents don’t cover something, it says so rather than making it up.

You build a base once and reuse it everywhere - one base can back many agents, and inside a flow individual steps can even use a different base than the rest of the call. This page walks through creating a base, filling it, testing it, attaching it, and how the look-up actually works on a live call.

Where it lives: the Knowledge base section in the left nav. A base does nothing on its own - it only affects calls once you attach it to an agent (covered below). Think of this section as your content library, and the agent as the worker who reads from it.

What a knowledge base is (and isn’t)

  • It is a searchable store of your own text - answers, policies, product facts, help-center pages.
  • It isn’t the agent’s personality or instructions. Who the agent is and how it talks lives in the agent’s Call context; the knowledge base is only the facts it can cite.
  • It isn’t a place for per-caller data like a name or an amount due - those are custom variables pulled from your contact list. A knowledge base holds facts that are the same for every caller (your policies, your catalogue), not values that change call to call.

Rule of thumb: if it’s true no matter who’s on the phone, it belongs in a knowledge base.

Create a base

From Knowledge base, hit Create base, give it a name, and you’re taken straight into it to add content.

Name

A label for you and your team - the caller never hears it. Name it after the content, not the campaign or the agent, so it’s still obvious later and so you can reuse it across agents.

  • Good: “Product FAQ”, “Refunds & returns policy”, “Loan eligibility rules”.
  • Weak: “KB for the September agent” - that tells future-you nothing about what’s inside.

Keep bases focused - one topic each. A tight “Refunds policy” base retrieves sharper, more relevant answers than one giant “Everything” base, because there’s less unrelated material to wade through. It’s better to have three small, on-topic bases than one sprawling one. (More on why under Retrieval below.)

A brand-new base is Empty until you add content. That’s the status you’ll see next to it - Empty means nothing’s indexed yet, Ready means it has searchable content the agent can pull from.

Add sources

Inside a base, the Add a source card gives you three ways to get content in. Each one is chunked and made searchable the moment you add it - no separate “publish” or “process” step, and no waiting. As soon as it says Ready, the agent can use it.

Text

Paste or type content directly. Give it a short title (e.g. “Refund policy”) and drop the content in the box below.

This is the best format for the answers that matter most, because you control the wording exactly. Write the questions customers actually ask, in the words they ask them, followed by the answer - like an FAQ.

  • Tip: short and direct beats long and flowery. “Refunds are processed within 5–7 business days to the original payment method” retrieves and reads back far better than three paragraphs of marketing copy around the same fact.

File

Drop a document in, or click to upload. As soon as you pick a file, it uploads and indexes - there’s no separate button to press.

  • Accepted formats: PDF, DOCX, TXT, MD, CSV, HTML, up to 10MB per file.
  • Great for content you already have written down - a policy PDF, a price list as a CSV, an exported help doc.
  • Tip: clean documents index better than messy ones. A PDF that’s mostly scanned images or dense tables gives the agent less usable text to work with than the same content as plain text. If an important file isn’t answering well, paste the key facts in as a Text snippet instead.

Paste a public web address and we crawl that page and index its text - handy for a pricing page, a docs page, or a help-center article you keep updated on your site.

  • Use a public https:// URL (pages behind a login can’t be crawled).
  • Add links one at a time - each link is one page. It doesn’t spider your whole site, so point it at the specific pages that hold answers.
  • Tip: a link is a snapshot taken when you add it. If you later change that page on your website, the base won’t know - re-add the link to pick up the new content (see Keeping it fresh).

What you’ll see after adding: each source shows up in the Sources list with its type and a chunk count. A “chunk” is just a bite-sized piece your content was split into for searching - more chunks simply means more material. The stat cards at the top tally your total sources and how many are text, files, and links.

Editing or replacing a source

Content changes. When it does:

  • Re-adding the same source replaces it - re-upload a file or re-add a link and the old version’s chunks are swapped for the new ones, so you never end up with two conflicting copies of the same document.
  • Removing a source deletes that source and all its chunks from the base. Deleting the whole base removes everything in it. Both are permanent, so the platform asks you to confirm.

Test it before you rely on it

Don’t attach a base to a live agent and hope. On a base’s page there’s a Test button that opens a chat panel where you can ask questions and see how the base answers - using the exact same look-up the agent uses on a call. It answers only from your sources, and tells you plainly when they don’t cover something.

This is the fastest way to find gaps. Ask it the ten questions your customers ask most:

  • If it answers well → great, that content’s landing.
  • If it says it doesn’t know → add a source for that question.
  • If it answers wrong → find the source that’s misleading it and fix the wording.

Do this before every launch, and again whenever you add or change sources.

Attach a base to an agent

A base only affects calls once an agent points at it. In the agent builder, on Step 1 (Basics), there’s a Knowledge base card:

  • Pick a base from the dropdown (each is listed with its chunk count so you can tell the full ones from the empty ones), or leave it on None for an agent that doesn’t need to look anything up.
  • One base can be attached to many agents - your “Refunds policy” base can back your support agent, your retention agent, and your win-back agent all at once. Update the base in one place and every agent that uses it gets the new answers.
  • Created the base in another tab? Hit Refresh on the card to pull it into the list without losing your half-filled agent form.

Attach it, then republish. Like any agent change, attaching (or swapping) a knowledge base only goes live on real calls once you Publish the agent. Test the base first, attach it, then publish.

Per-node knowledge base in a flow

If your agent uses the prompt & flow builder - where a call branches through steps like greet → qualify → pitch → book - each step (node) can decide which base to read from. Open a node’s inspector and you’ll find a Knowledge base field with three choices:

  • Inherit from agent (the default) - this step uses whatever base you set on the agent. Leave it here unless you have a reason not to; it means you set the base once and every step follows.

  • None - this step does no look-up at all, even if the agent has a base. Use it for steps that shouldn’t reference documents - a scripted greeting, a pure routing step, or the closing goodbye - so the agent stays on-script instead of pulling in facts it doesn’t need.

  • A specific base - this step reads from a different base than the rest of the call. Handy when one part of the call needs specialised material: a “pricing” step could pull from a pricing base while the rest of the call uses your general FAQ.

  • Tip: most flows never need this - set one base on the agent and let every step inherit it. Reach for per-node bases only when a step genuinely needs different (or no) source material.

How retrieval works

You don’t have to configure any of this, but knowing the shape of it helps you write better sources.

On every turn where the agent has a base, the platform:

  1. Takes what the caller just said.
  2. Searches the attached base for the passages most related in meaning to it - not just keyword matches, but semantic ones, so “can I get my money back?” still finds your “refund policy” content even though the words differ.
  3. Feeds the handful of most-relevant chunks into the agent’s thinking for that reply, so the answer is grounded in your content.
  4. Notes on the call transcript which sources it consulted, so you can see afterward where an answer came from.

Two things follow from this that are worth internalising:

  • Focused bases retrieve better. The search pulls the top few most-relevant pieces. In a tight, on-topic base those top pieces are almost always the right ones. In a sprawling everything-base, genuinely relevant content competes with unrelated material for those few slots - which is exactly why we keep recommending one topic per base.
  • The agent answers what’s in the base. It can only cite material you’ve added. If the base doesn’t cover a question, a well-briefed agent says it doesn’t know rather than inventing an answer - which is the safe, correct behaviour on a real call.

Keeping it fresh

The agent quotes whatever is in the base right now. So the base is only as trustworthy as your last update:

  • When a policy or price changes, update the source. Re-add the file or link (it replaces the old version) or edit the Text snippet. An out-of-date base means an agent confidently quoting last quarter’s prices.
  • Remember links are snapshots. Changing the page on your website does not update the base - re-add the link.
  • Re-test after changes. Pop open the Test panel and ask the questions that touch what you changed, to confirm the new answer comes back.

Quick tips

  • One topic per base. Sharper retrieval, easier upkeep. Split a giant base into a few focused ones.
  • Prefer short, direct answers over long marketing pages - write it the way you’d want the agent to say it.
  • Include the real questions customers ask, phrased the way they phrase them.
  • Keep prices and policies current - the agent will quote exactly what’s in the base.
  • Test before you trust, and again after every change.
  • Leave flow nodes on “Inherit” unless a step genuinely needs different or no source material.

Next: Tools & actions → - give your agent things it can do mid-call (transfer to a human, book a meeting, send a WhatsApp, push to your CRM), not just things it can say.