The Agentic Search pattern

In one line. Conversational, cited answers over your enterprise documents and data — a Knowledge collection plus an agent that reads it. You’ll be able to. Recognize when Search is the right shape, build one in four steps, and tune it from demo to dependable.

What it is

Agentic Search is level 1 of the solution ladder: the fastest way to turn a pile of documents into an expert your whole team can ask. Users type questions in plain language; the agent answers from — and only from — the content you attached, with citations that click through to the exact source passage.

Choose this shape when the sentence in your head is “let people ask questions over our content.” If the sentence is “automate this process,” you want a Workflow.

The pattern, in four steps

1. KNOWLEDGE collection  →  2. ADD documents  →  3. AGENT  →  4. ATTACH collection
   (purpose: Knowledge;      (upload or            (persona +     (Knowledge tab —
    indexed as it lands)      connector)            model)         the load-bearing wire)

Create a Knowledge collection — Studio ▸ Data ▸ Collections ▸ + Collection, then choose Knowledge under Purpose. That purpose is the fork in the road: Knowledge means no lifecycle — every document is indexed for your agents to read as soon as it lands, with no stages to walk. Give a big upload a few minutes.
Add documents — upload them directly, or connect a source under Studio ▸ Data ▸ Ingestion and schedule the pull so the collection stays current on its own (connectors). Is it ready? Open the collection’s Index Health tab. It shows how much of the collection is indexed and lists any documents that didn’t make it, so you can retry them there instead of blaming the agent later.
Build the agent — Studio ▸ Agents ▸ Agents ▸ + Agent. The persona sets the contract: answer from the attached documents, cite always, refuse when the answer isn’t there.
Attach the collection on the agent’s Knowledge tab. This attachment is both the power (grounded, cited answers) and the boundary (the agent cannot see anything you didn’t attach).

Your users never open Studio — they meet the finished assistant in the project’s chat and search channels, which you set up under Studio ▸ Experience ▸ Chat and Studio ▸ Experience ▸ Search (Chat & search).

Step-by-step with every screen: B1 · Policy assistant.

If you don’t see the Knowledge purpose when you create a collection, an administrator has switched it off for your workspace — ask them to turn it on.

Tuning it from demo to dependable

You want	Do this	Where
Trustworthy answers	Keep citations on; make the persona refuse out-of-scope questions	Agent Persona
No leakage	Attach only the collections this audience may see; build separate agents for separate audiences	Agent Knowledge tab
Structured questions (“how many policies mention X?”)	Add a Knowledge schema so facts also land in a Datasheet, which the agent can query directly	Collection schema + `Studio ▸ Data ▸ Datasheets`
A guided start for users	Configure suggested questions in the chat experience	Chat & search

The distinction worth remembering: reading questions (“what does the policy say?”) are answered from passages in your documents; counting questions (“how many contracts expire in Q3?”) are answered from the rows in a Datasheet. A great assistant does both.

Worked builds

Build	What it adds
B1 · Policy assistant	The whole pattern, smallest form — the build from your first 30 minutes, production-grade
B2 · Enterprise data Q&A	Multiple collections, Knowledge schema, structured + narrative answers together

Where to go next

Ready for a process, not just questions: the Workflow pattern.
The controls behind the agent: A3 · Your first agent.
How the collection purposes differ, and when to pick each: D2 · Collection types.