Skip to content

Chat with the AI about your code

Beyond one-shot descriptions, Understand has an AI Chat panel where you can ask follow-up questions and hold a conversation that stays grounded in the code you're looking at.

Chat quality and speed depend on what's running the model

Chat is only as good as the model behind it, and that's a choice you (or your admin) make, not something Understand fixes for you. A 1-billion-parameter model on a laptop and a 200-billion- parameter model on a data-center-grade server will give noticeably different experiences — both in how fast responses come back and how good the answers are. See choose a model for the size/quality/speed tradeoffs of the bundled options, or share one AI server across your team if you'd rather point everyone at a single, more capable machine instead of running small models locally.

Turn AI on first

Chat uses the same engine as the AI Overview. If you haven't enabled AI yet, see Understand AI — overview & setup.

Open AI Chat

Open it from View → AI Chat. The panel docks on the right. Its input box prompts "Ask about your codebase"; the toolbar gives you New Chat, Show History, Options, and Cancel.

Chat with entity context

When you open chat from an entity (or after viewing an entity in the AI Overview), the conversation carries that entity's context — so questions like "what does this function do?" or "what calls into this?" are answered about the thing you're actually looking at, and follow-ups build on the previous turns.

  • New Chat starts a fresh conversation.
  • Show History brings back earlier conversations.
  • Options jumps straight to the AI settings on the Data preferences page.

Chat history

Every chat is saved automatically — nothing to click to keep it. Show History lists past conversations for the project, most-recently-modified first; right-click one to Rename, Delete, or Export it to a JSON file. History is stored locally per machine, not in the shared project settings — it doesn't travel with the .und project if you hand it to a teammate or check it in.

Whole-project context is capped

Chatting from the project root's AI Overview grounds the conversation in a whole-project summary, which is capped at 50,000 characters — past that it's cut with a truncation notice rather than silently dropped or rejected. Chats scoped to a single entity or a smaller directory aren't affected by this particular limit.

  • AI Overview — the on-demand description of the current entity (see overview & setup).
  • AI Code Search (Search → AI Code Search) — semantic ("find code that does X") search built on a local embedding index. Enable it with the Enable Code Search checkbox on the Data page. The index itself lives locally alongside the project (not in the shared/portable project files), so it doesn't transfer with the project either — each machine builds its own.

Chat prompts include code context; for what stays on your machine (and what leaves it if you pick a cloud provider), see What can the AI access?. And if you'd rather chat in your own AI agent or editor, connect it to Understand via MCP — it gets the same exact code model.

If chat says AI is unavailable

If the assistant replies that AI is disabled or unsupported, the message states the reason — most often an unsupported platform (Intel Macs), a machine that failed the speed test, or AI simply turned off. See overview & setup.