The AI research assistant

What the assistant does

You type a question the way you would ask a colleague. Two kinds of question are what it is built for: questions about the Signal language and platform behaviour, which it answers from the documentation, and requests to compute something, which it runs and returns as a worked number or a chart.

It also works with the context of your open workspace. The assistant can see the signal or backtest in your active tab, so you can ask about your own work directly, not only about the platform in the abstract.

Grounded, cited answers

For any Signal-language question, the assistant searches the documentation before it answers. Syntax, a function, an accessor, sampling versus bucketing, null versus NaN, a worked example: it retrieves the relevant passages and answers from what it finds, not from memory. It does not invent DSL syntax, and it will not hand you a function name or an argument that is not in the reference.

Every grounded claim carries a small source pill inline, naming the documentation page and the section the passage came from. The pill sits at the end of the sentence it supports, so you can see which specific statement rests on which specific passage and open the source to read it in full. When several sentences draw on different pages, each gets its own pill.

The assistant also pulls likely-relevant documentation from your question before it starts writing, so a well-scoped question about, say, operators and functions tends to come back already grounded in the right pages.

It computes, not only chats

The assistant runs analysis in an isolated Python scratchpad and renders charts inline in the conversation. An answer can be a computed number, a table, or a plot, not only prose. Ask it to work through a payoff, fit a quick distribution, or draw a curve, and it writes the code, runs it, and shows you the result in line.

The scratchpad is a generic compute sandbox with a 60-second limit per run. It is not a Signal-language runner, and it cannot reach market data. It will not pull SPY's IV history or execute a backtest for you. It computes on numbers you give it in the conversation or on data it generates itself. Use it for the math and the visualization around a question, and keep the market-data work on the Signal and backtest surfaces where it belongs.

What it sees

The assistant reads two things: the documentation and the context of your open workspace. It can see the signal or backtest in your active tab, so if you ask why a backtest's drawdown looks the way it does, it has that backtest in view.

So it handles both kinds of question well: how the platform behaves in general (how an accessor works, why a window is observation-count, how to structure a Signal-language expression, or a calculation you hand it directly), and what is happening in the specific work you have open.

Conversation scope

A thread is one working session. The assistant remembers what you said earlier in the same conversation, so you can build on a previous answer without restating it. That history is scoped to the thread and is ephemeral: it is held in memory and does not persist across restarts. Starting a new thread starts fresh, with none of the earlier context carried over.

Keep one line of inquiry in one thread while it stays relevant, and open a new thread when you move to an unrelated question so the earlier context does not bleed in.

Trust note

The assistant treats documents and search results as reference material, not as instructions. If a passage it retrieves happens to contain text that reads like a command, it will not act on that text. Retrieved documentation informs the answer; it never redirects the assistant.