How to Connect Your Knowledge Sources (Notion, Slack, Zendesk) to an AI Support Agent

We'll use Macha as the worked example because it's the product we build and can show you real screenshots of, but the principles — sources vs. connectors, indexing, scoping, keeping things fresh — apply to any serious agent platform. One framing to set straight up front, because it shapes everything below: Macha is an AI agent layer that runs on top of your existing helpdesk (Zendesk or Freshdesk). It is not a helpdesk and not a Zendesk replacement — your tickets, your Help Center, and your agents stay exactly where they are. Macha adds an AI agent that reads your knowledge and works inside that setup. For the Zendesk-specific version, see Macha on Zendesk.

Why an AI agent is only as good as the knowledge it can reach

When a support bot gives a generic, useless answer, there's almost always one root cause: it was answering from its training data — the general internet it learned from — instead of your facts. Training data knows what a "refund" is in the abstract. It does not know that your refund window is 30 days, that opened software is non-refundable, or that EU orders route through a different warehouse. That gap is exactly where customers get burned and trust evaporates.

The fix is grounding: pointing the agent at your actual content and constraining it to answer from there. This is the same idea behind retrieval-augmented generation (RAG) — the agent searches a curated knowledge base, pulls the relevant passages, and writes its answer from those, not from memory. Done right, three things change:

• Answers get specific and correct — "Your return window is 30 days from delivery" instead of "Return policies vary."
• The agent can cite its source, so a human can verify it and customers trust it.
• The agent can say "I don't know" and escalate when nothing in your knowledge covers the question — which is far better than a confident wrong answer.

So the real work of deploying an AI agent isn't prompt-tuning. It's connecting it to good knowledge. Let's do that.

The key distinction: Sources vs. Connectors

In Macha, knowledge and actions are handled by two different things, and understanding the split makes the rest of the setup obvious.

• Sources are the knowledge your agent searches across to answer. Documents, websites, single web pages, and articles pulled from connected apps. This is the "what does the agent know?" layer. The Sources page literally describes itself as "Add knowledge sources for your AI to search across."
• Connectors are app connections that enable both knowledge ingestion AND actions. The Connectors page describes itself as "Connect your apps and services to enable AI-powered actions." A connector can feed a source (the Zendesk connector pulls in your Help Center articles) and give the agent tools to do things (read a ticket, post a reply, tag it, set status).

The overlap is the part people miss: some connectors double as knowledge sources. Connect Zendesk and you get both — your Help Center becomes a searchable source and the agent can act on tickets. Connect Notion or Confluence and those docs become live knowledge the agent reads. So the mental model is: Connectors are the pipes; Sources are the water. Some pipes (Zendesk, Notion, Confluence, Google Workspace) carry both knowledge and the ability to act; others (uploads, a crawled website, a single page) are pure knowledge.

With that straight, here's the end-to-end setup.

Step 1 — Connect Zendesk so the agent works on top of your helpdesk

Start with your helpdesk, because that's both where the tickets live and where a big chunk of your knowledge already exists (your Help Center). On the Connectors page, find Zendesk and click Connect. Zendesk authenticates via OAuth (click Connect, authorize through your Zendesk admin account) or an API key.

Connecting Zendesk does two jobs at once:

1. It ingests your Help Center as a knowledge source. Macha pulls your published articles in and keeps them current with a live webhook sync — when you publish or edit an article in Zendesk, the source updates. You'll see it appear on the Sources page as something like "Zendesk Help Center · 347 articles · via connector · Ready."
2. It gives the agent tools to act on tickets. The Zendesk connector exposes a full set of tools — read ticket details, search tickets, read image and file attachments, send public replies, add internal notes, update status/priority/tags and custom fields, search users and groups, and assign or route tickets. So the agent doesn't just answer; it can do the ticket housekeeping too.

This is the heart of the "layer on top" model: your Zendesk stays your system of record. Macha reads from it and writes back into it. If you haven't built out your Help Center yet, that's worth doing first — it's the highest-leverage knowledge you can give an agent. We have a full guide on how to build a Zendesk knowledge base, and a primer on how the Zendesk ticketing system works if you're newer to it.

Step 2 — Add your other knowledge sources

Your Help Center is rarely the whole picture. The answers to your trickiest tickets often live in an internal Notion wiki, a Confluence space, a policy PDF, or a single page on your marketing site. Add those on the Sources page.

At the top of the Sources page you get three direct add options:

• Upload Documents — drop in static files: PDF, DOCX, TXT, CSV, XLSX. Good for policy docs, manuals, FAQs, and structured data. The limits: up to 10 files at a time, 20 MB each. (Handy detail: small spreadsheets of 2,000 rows or fewer get injected directly into the agent's context so it always has the full table; bigger files get chunked and made searchable.)
• Add Website — crawl and index an entire site. Macha follows links, discovers pages, and indexes them, up to a system limit of 200 pages. Use it to pull in a whole documentation site or public help center.
• Add Webpage — index a single URL. Perfect when you only need one specific page — a pricing page, one detailed FAQ — without crawling everything around it.

Then there are the connector-based sources, which is where Notion, Confluence, and Google Workspace come in:

• Notion — connect via OAuth and select which pages and databases to share. Macha reads them live using its Notion tools, so when someone edits the doc, the agent sees the new version. (If you're coming from a Zendesk angle, we cover the broader picture in how to connect Notion with Zendesk.)
• Confluence — connect with an API token and the agent can search and read your spaces. One gotcha worth flagging: Confluence API tokens expire, and Macha will warn you ~30 days before so the source doesn't silently go dark.
• Google Workspace — connect via OAuth and pick specific Docs and Sheets; the agent reads them live.

A nice convenience here: when you add a connector-based source (Notion or Google) to an agent, Macha auto-attaches the read tools the agent needs — you don't have to wire them up by hand.

A word on Slack, since it's in the title. Connect Slack (OAuth) and the agent can operate in Slack — responding when mentioned in a channel and handling DMs to the Macha app. That makes Slack a useful action surface and internal channel for the agent. Be honest about the distinction, though: for document knowledge — the policies and how-tos the agent answers from — your workhorse sources are Notion, Confluence, Google Docs, uploads, your website, and your Zendesk Help Center. Slack is where the agent can show up and answer, more than a library of documents it indexes.

Step 3 — Let Macha index your sources (the "Ready" state)

Connecting a source isn't instant magic — there's a processing step, and it's worth understanding so you don't test too early and conclude "it doesn't work."

After you add a source, Macha indexes it: it reads the content, splits it into chunks, and creates embeddings (vector representations) so the agent can search it semantically. Uploaded files usually finish in seconds to a couple of minutes; a website crawl takes a few minutes depending on size. Each source carries a status badge so you know where it stands:

• Ready — the source has been indexed and the agent can search it. This is what you'll see on uploaded files, indexed websites, and your Zendesk Help Center.
• Live — the source is read in real time directly from a connected app (Google Workspace, Confluence, Notion), so the agent always reads the current version rather than an indexed snapshot.

The practical takeaway: wait for Ready (or confirm a Live source is connected) before you start testing. A common false alarm is "the agent ignored my doc" when the doc simply hadn't finished indexing yet.

One honest note on cost while we're here: the indexing and embedding step is not charged against your credits. Macha's credits are spent per AI action — any automated step the agent takes, like searching, drafting a reply, tagging, or resolving — not for ingesting your knowledge and not per resolved ticket. (More on the per-action model below.)

Step 4 — Build an agent and point it at those sources

Knowledge sitting in the library does nothing until an agent is told to use it. Create or open an agent and assign the sources it should search. This is also where you decide scope — which is one of the most underrated levers for answer quality.

Each source you add to an agent has an inline scope toggle:

• All documents — the agent can use every document in that source, and automatically gains access to new ones as they're added.
• Selected — the agent is restricted to the specific documents you pick.

Why scope matters: a billing agent shouldn't be wading through your engineering runbooks, and a tier-1 deflection agent probably shouldn't see internal escalation playbooks. Narrower, relevant knowledge produces sharper answers and fewer weird tangents. If you assign a Notion or Google source, remember Macha auto-adds the matching read tools, so the agent can actually fetch the live content.

You can mix source types freely on one agent — a small pricing spreadsheet injected directly, a big policy PDF as searchable knowledge, and a live Notion page — and the agent will pull from whichever fits the question.

Step 5 — Test that it answers from YOUR knowledge (and escalates when unsure)

Now the part most people skip and then regret. Don't trust it blind — interrogate it.

1. Ask a question your docs clearly answer. Something specific to your business: "What's the return window for opened items?" The agent should give your exact policy, not a generic one. If you can, check that it references the source it used.
2. Ask something only one specific source covers. This confirms the right source is wired up and in scope — e.g. a detail that only lives in a Notion page.
3. Ask something nothing covers. This is the most important test. A good agent should not invent an answer. It should say it doesn't know and escalate to a human — handing the ticket off with context — rather than hallucinate. If it makes something up, that's a signal your knowledge has a gap or the agent's instructions aren't strict enough about staying grounded.
4. Check a borderline case. Vague or multi-part questions reveal whether the agent retrieves the right passages or grabs something adjacent-but-wrong.

If an answer is wrong, resist the urge to immediately rewrite the prompt. Nine times out of ten the real fix is in the knowledge: a missing doc, an out-of-date article, two sources that contradict each other, or a source that wasn't in scope. Fix the content first.

Step 6 — Keep your sources fresh

Knowledge rots. A policy changes, a price moves, a product gets renamed — and an agent confidently citing last quarter's facts is worse than no agent. Build a refresh habit:

• Uploaded files don't auto-update. If you change the original document, Macha won't notice. You have to delete the old file and re-upload the new version. For anything that changes often, prefer a live connector source (Notion, Confluence, Google Docs) so the agent always reads the current copy.
• Websites need a manual Resync. Macha doesn't watch your site for changes. Hit the Resync button on a website source to pull the latest pages — a resync also clears out pages that have been removed.
• Zendesk Help Center stays current automatically via the live webhook sync, so editing an article in Zendesk flows through.
• Watch token expiry. Confluence API tokens expire; renew them when Macha warns you so that source doesn't quietly stop returning results.

Best practices (and the garbage-in, garbage-out reality)

Connecting sources is easy. Getting good answers is a content discipline. A few rules that consistently separate agents that help from agents that embarrass you:

• Clean docs win. A well-structured, current help article beats ten messy ones. If your own team can't find the answer in your docs, neither can the agent. Tidy the source, don't just bolt on more sources.
• Scope sources per agent. Give each agent only the knowledge its job needs. Broad, unscoped access is the most common cause of confident-but-irrelevant answers.
• Retire stale pages. Out-of-date content is actively harmful — the agent can't tell "old but published" from "current." Archive or delete superseded articles; resync after you do.
• Resolve contradictions. If two sources disagree (an old PDF says 14 days, the Help Center says 30), the agent may pick either. Single source of truth per fact.
• Garbage in, garbage out is real. No model overcomes bad knowledge. "Connecting" a source is not the same as the source being good. Most "the AI gives wrong answers" problems are content problems wearing an AI costume.
• Test the escalation path, not just the happy path. The greeting always works. Whether the agent honestly hands off what it can't answer is what actually protects your customers.

Where this fits — and an honest cost note

The whole point of grounding an agent in your knowledge is that it stops deflecting with a vague article link and starts resolving — pulling the real answer from your Help Center, Notion, or Confluence and replying inside the ticket, while still tagging, routing, and escalating with context when it isn't confident. That's the layer Macha adds on top of Zendesk or Freshdesk — not a replacement for them.

The honest framing on cost: Macha bills per AI action — any automated step the agent takes (searching your knowledge, drafting a reply, tagging, routing, resolving) — not per resolution or per ticket, because most of the useful work an agent does isn't a tidy "resolution," it's the steps along the way. Indexing your knowledge doesn't burn credits. And the agent is only ever as good as what you connect it to — which is the entire reason this guide exists. You can try it free — 7-day free trial, no credit card required.

Frequently asked questions

What's the difference between a source and a connector? A source is knowledge your agent searches to answer — documents, websites, single pages, and articles from connected apps. A connector is an app connection that enables both knowledge ingestion and actions. Some connectors do both: the Zendesk connector pulls your Help Center in as a source and lets the agent read and update tickets; Notion and Confluence connectors turn those docs into knowledge the agent reads.

Can I connect Notion to an AI agent? Yes. Connect Notion via OAuth and pick which pages and databases to share. Macha reads them live, so the agent always sees the current version, and it auto-attaches the Notion read tools to the agent when you add the source. See how to connect Notion with Zendesk for the broader picture.

What knowledge formats can I upload? Direct uploads support PDF, DOCX, TXT, CSV, and XLSX, up to 10 files at a time and 20 MB each. For web content you can crawl a whole site (up to 200 pages) or index a single URL. For frequently changing content, use a live connector source (Notion, Confluence, Google Docs) instead of static uploads.

How long does indexing take, and how do I know it's done? Uploaded files usually index in seconds to a couple of minutes; a website crawl takes a few minutes. Watch the status badge: Ready means the source is indexed and searchable; Live means the agent reads it in real time from the connected app. Wait for one of those before testing.

Does connecting a source guarantee good answers? No — and this is the honest part. An agent is only as good as the content behind it. If your docs are stale, contradictory, or thin, the answers will be too. Connecting a source is step one; keeping that source clean, current, and correctly scoped is what actually drives answer quality. Garbage in, garbage out applies fully.

Will my Zendesk Help Center stay up to date in the agent? Yes. The Zendesk connector keeps your Help Center source current with a live webhook sync, so publishing or editing an article in Zendesk flows through automatically. Uploaded files and crawled websites, by contrast, need a manual re-upload or Resync to refresh.

Does indexing my knowledge cost credits? No. Indexing and embedding aren't charged against credits. Macha's credits are spent per AI action the agent takes — searching, drafting, tagging, resolving — not for ingesting knowledge and not per resolved ticket.

The bottom line

An AI support agent is a search-and-reasoning engine bolted onto your knowledge — so the quality you get out is set by the quality you connect in. The sequence is straightforward: connect Zendesk (it ingests your Help Center and lets the agent act on tickets), add your other sources (Notion, Confluence, Google Docs, uploaded files, a crawled website, or a single page), let them reach the Ready/Live state, assign and scope them to an agent, test that it answers from your content and escalates when it can't, then keep everything fresh. Remember the two-part mental model — Connectors are the pipes, Sources are the water, and some pipes carry both knowledge and the ability to act. Get the knowledge right and the "AI" part mostly takes care of itself. Skip it, and no model on earth will save you. For the Zendesk-specific setup, see Macha on Zendesk.

Product details verified June 2026 against Macha's own documentation and dashboard. Macha updates its product periodically — confirm labels and limits in your own account before relying on them.