Welcome!

AI Agent per Library

TL;DR; Each library can have its own AI agent — a chat experience that answers questions using only the content items in that library. It's one of Lupo's most distinctive features and is very close to production quality. If your plan includes it, turn it on.

Courses are good for sequenced learning. A browsable knowledge base is good for on-demand reference. An AI agent on top of that knowledge base is a third thing: a conversation that gives you a direct answer to a specific question, grounded in your organization's content.

What the AI agent is

The AI agent is a chat interface scoped to a single library. A learner types a question — "how long is parental leave?", "what's our incident response process?", "which tooling do we use for code reviews?" — and the agent answers using the content items in that library as its source material.

A few things that make it specifically useful:

  • It only uses the library's content. The agent doesn't go to the broader internet and doesn't pull from other libraries the learner doesn't have access to. Its answers are grounded in the content items you've approved.
  • It cites what it reads. Answers typically reference the specific content items they came from, so the learner can click through and verify.
  • It stays inside the permission model. If a learner doesn't have access to a library, the agent doesn't answer from it — see Library Visibility.

When it's worth turning on

Enable the AI agent on a library when:

  • The library has enough content to be useful. An agent pointed at three PDFs isn't much better than reading the three PDFs directly. The value kicks in once there are dozens of documents and the learner doesn't know where to start.
  • Learners ask the same questions repeatedly. If your HR team or Support team answers "how does PTO work?" five times a week, that's exactly what the agent is for.
  • The content is reasonably well-written and well-titled. The agent is smart, but it still benefits from clear titles and descriptions on the content items. See Uploading Content Items for tips.

Don't enable the agent on a library that's still being built out. Get the content in place first, make sure titles and descriptions are clean, then turn on the agent.

Enabling the agent

From the library's settings, look for an AI Agent or Chat section and turn it on. Depending on your plan, there may be options for:

  • Which content items are indexed — usually "all of them" by default, but you can exclude specific items if some content shouldn't be answerable.
  • A system prompt or persona — a short paragraph telling the agent how to respond, what tone to use, and any hard rules (e.g., "always remind the learner to check the date before acting on compliance info").
  • A default first message — what the agent says when someone opens the chat without typing anything.

Once it's enabled, the agent shows up as a chat pane or button on the library page. Learners can start asking questions immediately.

"Almost at production quality"

The AI agent is a genuinely great feature, and it's close enough to production that you should feel comfortable rolling it out. A few things are worth knowing as you do:

  • It's still evolving. Response quality, citation behavior, and the exact UI can change as the feature matures. Don't embed it so deeply in a critical workflow that a minor change breaks you.
  • It's not a replacement for course-level tracking. The agent helps people find answers; it doesn't track whether they learned anything. For compliance training or formal onboarding, still use a course with activities and progress tracking.
  • It respects the data it's given. If the library has wrong or outdated information, the agent will confidently repeat it. Keep the content current — the AI makes bad content more visible, not less.

Tips for getting good answers

The biggest lever is the quality of your content, but there are a few other things that help:

  • Put clear titles and descriptions on every content item. Even though the AI indexes the full document, the title and description are the strongest signal about what the document is actually about.
  • Break long documents into smaller, focused files. A 200-page handbook with 40 topics is harder for the agent to answer from than 40 separate 5-page documents.
  • Use consistent terminology. If half your docs say "PTO" and half say "vacation days," the agent can still handle it, but learner questions line up with your content better when the vocabulary matches.
  • Review the agent's answers yourself. Before rolling it out to the whole department, spend 20 minutes asking it the questions you know learners will ask. You'll find gaps in the content that are easy to fix before anyone else sees them.

What the agent doesn't do

A few deliberate limitations worth knowing about:

  • It doesn't perform actions — it won't update documents, file tickets, or send emails.
  • It doesn't remember past conversations across sessions (for privacy reasons).
  • It doesn't promise certainty. If a question has no good answer in the library, the agent says so rather than guessing.

If you need the agent to do any of those things, you're outside what this feature is for today.

Where to go next