Training Murphy on your content

Murphy only answers from what you give it. That is the point: a support agent that invents policy is worse than no agent.

Sources

Open Training on your agent. Four kinds of source:

Links. Give Murphy a list of URLs and it crawls them. Best for docs sites, help centers, and pricing pages. Re-crawl when the content changes.

Files. Upload PDFs and documents. Murphy reads PDFs directly, including ones customers send mid-conversation.

Text. Paste anything. Useful for policies that live in someone's head, like the actual refund rule rather than the published one.

Q&A pairs. Write the question and the answer yourself. This is the sharpest tool you have. When Murphy gets something wrong repeatedly, a Q&A pair fixes it faster than rewriting a page.

How retrieval works

Content is split into sections, embedded, and stored per agent. On each question Murphy retrieves the closest sections and answers from those, carrying the source through so you can see what it read.

Sections are tagged with their document title, so an answer drawn from your refund policy knows it came from the refund policy. That provenance shows up in the inbox next to the answer.

Docs pushed by your coding agent

If your knowledge lives in a codebase rather than on a website, your own coding agent can write it. POST /api/v1/knowledge takes a title, a slug, and markdown content, embeds it, and replaces any previous version at the same slug.

This is how Murphy's own docs work. The page you are reading is a knowledge doc in Murphy's workspace, which is also what Murphy's support agent answers from. One article, two surfaces, no sync step to forget.

See the API reference for the full shape.

When Murphy does not know

It says so, and offers a handoff rather than filling the gap with something plausible. Those moments are logged. Monitors scores conversations nightly and builds a review queue, so the honest gaps become a to-do list instead of disappearing.