PerchAI
How it works

How Perch works

Perch reads your material first, grounds its answer in what it found, and shows the source behind the claims it makes from it.

Perch follows one principle on every turn: read first, then answer, and keep the source attached. This page explains the loop behind that principle so you know what is happening when you ask Perch to do something.

Read firstGather the documents relevant to the request.
Do the workRead, search, and run analysis in a sealed sandbox.
Keep the source attachedCite each claim; mark anything unsupported.

Read first

When you ask a question, Perch first gathers the material that is relevant to your request:

  • The documents in your workspace, whether those are uploaded files or local folders.
  • What Perch has learned about you and your work over time.
  • The public web, when the question needs current or outside facts. See Web search.

This step is called retrieval. It is the reason Perch can answer about your March invoices or your project folder rather than about the world in general.

When nothing relevant is found, Perch can still answer stable, general facts from what the model knows, without a citation. For anything current, it searches rather than answering from memory.

Do the work

Once Perch has the relevant material, it does the task. The shape of the work depends on the surface:

  • On the web, a turn is a single grounded answer with its sources attached.
  • On the desktop and CLI, Perch acts as an operator. It can read files, search across a folder, and run analysis in a sealed sandbox to check numbers and produce results rather than estimating them.

For larger tasks, Perch can plan a sequence of steps and use focused workers for parts of the job. The driving model is chosen at the start of a task and stays fixed for that task, so the work stays consistent from beginning to end.

How much Perch does on its own is up to you. The composer's Ask, Plan, and Agents modes decide whether a turn only answers, drafts a plan for your approval, or does the work. A goal keeps Perch working across many turns until the objective is met.

Keep the source attached

The output is only half the job. When Perch answers from your documents or from web results, it cites the source beside the claim so you can check it rather than trust it. A claim the material does not support is said plainly or marked as unverified, not dressed in a citation.

You can follow a cited claim back to the material it came from. The mechanics of that are covered in Citation verification.

One account across surfaces

The web chat, the desktop app, the CLI, and Perch on your phone are one product. They share:

  • One account. Sign in once; your access carries across surfaces.
  • One memory. What Perch learns on one surface is available on the others. See Memory.
  • One usage allowance. A single allowance covers every surface.

The result is that you can start a question in the browser and continue the work in the desktop app without re-explaining yourself.