Internal Knowledge Base Search for Support Agents: Find the Answer Inside the Ticket

Internal knowledge base search for support agents fixes the findability problem, not the documentation problem. You already wrote the docs. What you're missing is a way for an agent to ask one plain-English question and get the exact internal answer surfaced right where they're working — without leaving the ticket. With Macha, an agent types the question, Macha searches across your Google Docs and Notion, finds the relevant pages, and drops the key excerpts straight into the conversation. No hunting, no guessing, no stale bookmark.

A quick framing note, because it matters: Macha is not a wiki and not a helpdesk. It's the AI agent layer that sits on top of the helpdesk you already use — Zendesk, Freshdesk, Gorgias, or Front — and connects your knowledge and commerce tools to it. So this isn't "move your docs into another tool." It's "let an agent search the docs you already have, from inside the tool they already live in."

What this actually looks like

Here's the canonical flow. An agent is working a ticket about response-time commitments and needs the real numbers. Instead of going to find them, they ask Macha:

Agent: "Can you find our SLA terms for enterprise customers? I need to respond to a ticket about response times."

Macha runs two searches in parallel — one against Google Docs, one against Notion — reads the most relevant hits, and replies inline:

Macha: Here's what I found: Enterprise SLA (from Google Docs — "Enterprise SLA Agreement v3"): • Critical issues: 1-hour response, 4-hour resolution • High priority: 4-hour response, 24-hour resolution • Normal: 24-hour response, 72-hour resolution Support Tiers (from Notion — "Enterprise Support Tiers"): • Enterprise customers get a dedicated Slack channel • Priority routing — tickets skip the general queue Want me to draft a reply using these details?

The agent reads two excerpts, confirms they're right, and is back to writing — with a citation to the source doc so they can verify, not just trust. That last line matters too: because Macha is an action-taking agent, the natural next step ("draft the reply") is one message away. But the agent stays in control: nothing is sent without their say-so.

The screenshot above shows the Sidekick surface where this happens; the specific SLA exchange below is the flow we capture next.

Why "search" beats a static knowledge base

A traditional internal knowledge base assumes the agent knows where the answer lives and what it's called. Real life isn't like that. The doc is titled "Enterprise SLA Agreement v3" but the agent searched "response time guarantee." The Notion page says "client onboarding" but the agent typed "customer setup." Keyword search misses all of it.

Macha's lookup is semantic — it matches on meaning, not on the exact string — so the agent can ask the way they'd ask a senior teammate. The distinction is well established: Elastic and Redis both describe traditional keyword (lexical) search as literal token matching — ranking documents by scoring methods like TF-IDF or BM25 — while semantic search compares vector embeddings, numerical representations of meaning, so a query for "response time guarantee" can surface a doc titled "Enterprise SLA Agreement." Because Macha reads the documents and returns excerpts rather than just a list of blue links, the agent gets the answer, not another search results page to triage. Findability is the product.

And the payoff is measurable, not hand-wavy. In the largest field study of its kind, NBER researchers Brynjolfsson, Li and Raymond found that giving customer-support agents a generative-AI assistant raised productivity by about 14% on average — and up to 34% for newer, less-experienced agents. Industry agent-assist deployments report similar gains, commonly 15–30% reductions in average handle time, driven mostly by killing the context-switching this use case targets. Surfacing the right internal answer at the moment an agent needs it is precisely where that time comes from.

The two ways Macha can reach your docs

This is the part most write-ups skip, and it's the part that actually determines accuracy and cost. Macha can connect the same Google Docs and Notion content in two distinct ways, and good setups often use both.

1. As a synced Knowledge Source (the RAG path)

You connect Google Workspace or Notion as a live Knowledge Source. Macha indexes the content into its embeddings pipeline and keeps it in sync as the docs change — so when an agent asks a question, retrieval is fast, semantic, and grounded in whatever the doc says today, not a snapshot from last quarter.

This is the right default for stable reference material an agent asks about constantly — SLA terms, policy language, troubleshooting runbooks, product specs. It's pre-indexed, so lookups are cheap and instant.

2. As live tool calls (the on-demand path)

The second path is direct search at request time. Notion exposes Search Pages and Get Page tools; Google Workspace exposes Search Docs and Get Document. When you give an agent these tools, it can run a fresh search against your live workspace mid-conversation — pulling the newest content even if it was published five minutes ago.

This is the right choice for fast-moving or long-tail content — a wiki that changes daily, or a question so specific it was never going to be pre-indexed usefully. The trade-off: a live call costs a little more than hitting a pre-built index, and it's only as fast as the upstream API.

The practical rule of thumb: sync your evergreen reference docs as a Knowledge Source for speed and cost, and grant live Search/Get tools for the long tail and the freshly-published. Most teams end up running both, which is exactly what the use case spec assumes — Google Docs and Notion, indexed and on-demand.

Setting it up

You don't need to be technical to wire this together. The shape is always the same:

1. Connect your helpdesk. Macha attaches to Zendesk, Freshdesk, Gorgias, or Front — the place your agents already work. If you're on Zendesk, the Macha on Zendesk page walks through the install.
2. Connect your doc stores. Authorize Google Workspace and Notion. For the synced path, add them as Knowledge Sources; for the on-demand path, enable their search tools.
3. Build the agent. Create an agent whose job is internal lookup, give it the knowledge sources and/or the Search/Get tools, and tell it in plain language how to behave — cite the source doc, return excerpts not summaries, ask before drafting a reply.
4. Decide the trigger. This use case is most natural as on-demand assist — the agent invokes it whenever they're stuck (via Macha's agent-facing chat). You can also wire it to a Ticket Updated event so Macha proactively searches when an agent flags that they need internal information. Both work; on-demand keeps the agent firmly in the driver's seat.

Full configuration steps live in the docs, and there are related, ready-to-borrow patterns in the use case library — including drafting full replies using Google Docs context once the lookup half is solid.

Where this fits — and where it doesn't

Agent-assist lookup is the conservative, high-trust end of the AI-support spectrum, and that's a feature. A human reads every answer before it reaches a customer, so the failure mode isn't "wrong reply sent" — it's "agent has to double-check," which they were going to do anyway. Use it when:

• Your answers live in internal docs that customers should never see verbatim (pricing logic, SLA terms, escalation rules).
• Agents handle varied, non-scripted questions where a canned macro won't cut it.
• You want the speed of automation with a human gate — the agent stays accountable for what goes out.

Honest watch-outs

This isn't magic, and pretending otherwise helps no one.

• Garbage in, garbage out. Macha surfaces what your docs say. If "Enterprise SLA v3" contradicts "Enterprise SLA v2" and you never archived v2, an agent can get the wrong number with full confidence. The fix is the boring one: keep an owner and a review date on your reference docs. This use case rewards tidy documentation and punishes a messy one.
• Permissions are yours to scope. Macha searches the workspaces you connect. Be deliberate about which Google Drive folders and Notion spaces an internal-lookup agent can reach — you don't want a customer-facing agent able to surface board-deck commentary. Connect the support knowledge, not the whole company drive.
• It's assist, not auto-pilot. By design, this pattern keeps a human in the loop. If your goal is fully deflecting tickets with no agent involved, that's a different agent (auto-resolve), with a different risk posture. Don't expect zero-touch from a workflow built to put an answer in front of a person.
• Lookups cost credits. Every search and every reply is an AI action. In Macha, that's credit-based — 0.5 to 9 credits per action depending on the model, with the default GPT-5.4 Mini at 1 credit. Pre-indexed Knowledge Source lookups are the cheap path; live tool calls cost a touch more. None of it is free, but the math is usually trivial next to the ninety seconds of agent time it saves. See the pricing page for how credits map to plans.

A note on accuracy and trust

The single most important habit to build into this agent: make it cite. When Macha returns "1-hour response, 4-hour resolution," the reply should name the source — "from Enterprise SLA Agreement v3" — so the agent can click through and confirm in two seconds. An answer with a citation is a tool; an answer without one is a guess the agent now has to go verify from scratch, which defeats the point. Macha returns the source by default in this pattern; keep it that way.

FAQ

Which doc tools can Macha search? Out of the box, Google Workspace (Google Docs) and Notion — both as a synced, always-current Knowledge Source and via live Search / Get tool calls. If your runbooks live in Confluence, there's a sibling pattern for that too.

Does this replace our knowledge base? No. Macha sits on top of the docs you already keep in Google Docs and Notion and on top of the helpdesk you already use. It doesn't ask you to migrate anything — it makes what you have findable from inside the ticket.

Will the agent send replies automatically? Not in this pattern. It surfaces internal information for a human agent and offers to draft a reply, but the agent confirms before anything is sent. If you want hands-off deflection, that's a separate auto-resolve agent you'd configure deliberately.

How fresh are the answers? A synced Knowledge Source re-indexes as your docs change, so answers reflect the current version. For content that changes by the hour, grant the live Search Pages / Search Docs tools so the agent reads the workspace in real time.

What does it cost to run? It's credit-based — each AI action is 0.5–9 credits by model (default GPT-5.4 Mini = 1). Pre-indexed lookups are the cheapest path; live tool calls cost slightly more. There are no per-seat surprises; see pricing.

Can I limit what it can see? Yes — you choose which Google Drive folders and Notion spaces the agent can search. Scope it to your support knowledge, not the entire company workspace.

Try it

If your agents lose real time hunting for answers that already exist somewhere in Google Docs or Notion, this is one of the fastest wins Macha offers — low risk, human-in-the-loop, and live in an afternoon. Start a 7-day free trial, no credit card required, connect your helpdesk and your docs, and let an agent ask the first question. Or read the docs for the full walkthrough.

Written by Abbas (Customer Support & AI, Macha) · Reviewed by Ankeet Guha (Co-founder & CTO) · Published 2026-06-24 · Last updated 2026-06-24.