Skip to main content
When you ask a question, the agent searches your indexes to find the documents that contain the answer. Most of that work happens automatically in the background, but what you see in the chat, and what you ask the agent, both shape how well it lands on the right material. This page covers what search looks like in the chat UI and how to write prompts that help the agent find what you need.

​
What search looks like in a chat

After you send a message, the chat shows a few things while the agent works:
  • A thinking indicator animates below your message. Short progress text describes what the agent is doing (“Searching your index…” and similar). The indicator disappears when the agent finishes composing its reply.
  • The agent’s response streams in as it’s ready. For straightforward questions, the answer appears quickly; for questions that require several searches or reading through long documents, the response takes longer.
  • When the response is complete and the agent drew from documents, a References button appears in the message’s actions. Click it to open the References panel on the right side of the chat.

​
The References panel

The References panel is where you confirm the answer against the source material. It opens beside the chat and shows one card per document the agent used:
  • Document name and display name identify the source.
  • A location indicator tells you where in the document the relevant content is. This looks like Page 3 for PDFs, Sheet Foo for spreadsheets, Lines 42-85 for text files, or a Section heading for Word documents and HTML pages.
  • A snippet shows the passage the agent actually used, so you can verify it without opening the full file.
Hovering a reference card highlights the corresponding citation in the agent’s response, making it easy to see which part of the answer came from which source. Clicking a reference opens the document viewer at the exact page or section the agent cited. If the agent answered without drawing from your documents (for example, it said “I don’t have that information” or it clarified its own limitations), the References button doesn’t appear. A missing References button is itself a signal that the answer isn’t grounded in your indexed material.

​
Helping the agent find the right documents

The agent’s search is only as good as the question it’s searching on. A few habits consistently produce better retrievals.

​
Include the terms that appear in your documents

If your product manuals always call a part “flow sensor assembly,” searching for “water meter thingy” is a bad match. Use the vocabulary your documents actually use: part numbers, model names, product lines, official component names. If you’re not sure, start with a broader term and ask the agent to narrow down:
What do we call the component on the XYZ-500 that measures incoming water flow?
Once you have the canonical name, ask your real question.

​
Give context that disambiguates

“What’s the pressure rating?” applies to a thousand different parts. “What’s the inlet pressure rating on the XY-4400-HD?” finds one. Every extra scoping detail (model, revision, region, firmware version) helps the agent pick the right document instead of averaging across many.

​
Ask one question at a time

Agents handle focused questions better than compound ones. Asking “what are the specs, common failure modes, and replacement parts for the XYZ-500?” forces the agent to spread its search budget across three loosely related topics. Break it up; the chat is designed for follow-ups.

​
Name the document if you know it

If you already know where the answer lives, tell the agent:
In the XY-4400-HD service manual, which step comes after draining the tank?
This turns an open-ended search into a targeted read and usually returns the answer faster and more accurately.

​
Rephrase, don’t repeat

If the first search misses, don’t re-send the same question. Tell the agent what was wrong or what you’re actually looking for:
That answer is about the 400 series. I need the 500. Try again.
You cited the installation guide; the answer should be in the service manual. Can you search there?
The agent reads your correction and adjusts.

​
When search doesn’t find what you expect

If the answer isn’t in the response or the References panel looks off, try these in order:
  1. Check which index the agent is searching. The chat header shows the agent’s name; the agent’s configured indexes determine what it can see. If your document is in a different index than what this agent uses, the agent can’t reach it. An admin can expand the agent’s reference scope if needed.
  2. Open a reference and verify the excerpt. If the agent cited the right document but extracted a misleading snippet, push back: “That passage is about the 2022 spec, not the current one. Can you find the updated section?”
  3. Ask the agent what it searched. “Which documents did you look at for that answer?” or “What keywords did you search for?” often surfaces why the match was off, and gives you something concrete to correct.
  4. Rephrase using the document’s own language. Open a related document, note the terms it uses, and try again.
  5. Report it. If a question that should have a clear answer keeps missing, use the thumbs-down button and explain what went wrong. See providing feedback.