Ask your system
A read-only chat over the System Map — what it can answer, how every answer cites the nodes it stands on, and why it says the map has nothing rather than guess.
Ask your system is the Ask tab on both System Map pages. You ask about the system — what calls the billing service?, which endpoints reach the customer table?, what changed in the last map version?, what would touching the payments module affect? — and it answers from the map, citing the nodes it found. With a repository chosen in the page's scope picker, the conversation is about that repository; a search that finds nothing there widens to the whole organization and says so.
It is the same tool loop as the concierge, with one difference: nothing it can call writes anything, so it never asks you to confirm anything. The only rows a question adds are the conversation itself.
What Ask does not do yet is listed in the implementation status.
What it can do
The model has eight tools, all reads over the map and the knowledge you may already read:
| Tool | Answers |
|---|---|
| Find nodes | Nodes by name, key or path fragment, optionally of one kind |
| Neighbours | The edges around a node, in, out or both, up to two hops, with counts by edge kind |
| Impact of | For up to ten nodes: dependents, downstream services, outside systems, endpoints, data entities and their flags, covering tests and incidents, each as a count with the nodes behind it, and the impact level those counts reach |
| Path between | The shortest chain of edges between two nodes, either direction, up to six hops — or that there is none |
| Search knowledge | Cited passages from the knowledge sources you may read, and from the repositories' own documents on the map |
| Read docs | The repositories' generated docs: a search across them, or one page or section, each cited D<page>:<section> |
| Recent changes | Which nodes and edges the latest map versions added, changed or removed, with each version's status |
| Runtime status | What production says about a node: for a runtime, which commit each environment is running and since when; for a service or module, the runtimes that run it; for an incident, its severity, status, link and the deploy running when it started; and the known production incidents on it |
Every tool call is shown in the conversation as a line, the way the concierge shows its work, so you can see what was looked up before the sentence that follows it.
The level is the impact analysis's rule. Impact of computes its level with the same written thresholds as a card's impact analysis, over the counts of what you may see. The model is told never to state a level a tool did not return.
Every answer cites its nodes
An answer names what it stands on in the forms the tools returned: N<id> for a node, E<id> for an edge, K<source>:<chunk> for a knowledge passage and D<page>:<section> for a section of the generated docs. They are shown as chips under the answer, and a node chip opens the same node sheet as the map's lists, so a claim can be checked in one click.
The platform, not the model, decides whether an answer is grounded. It reads the citations out of the answer and matches them against what this question's tools actually returned. A citation no tool returned is named under the answer as not from the map, and an answer that cites nothing from the map is marked that way and shown as conversation, not as a fact. Under a grounded answer, a line says how many citations it rests on.
When the map has nothing
When the tools the model tried all come back empty — no node matches, no passage is relevant — the model's prose is not shown at all. The platform writes the answer itself from what was searched for, headed Not on the map: The map has no node for "invoicing". It does not answer from general knowledge of how systems like yours are usually built, and it does not guess a dependency the map does not record.
The chat reads only what the analysis indexed. On a limited map an absence is not proof: check the build status on the page before reading "none" as "none exists".
What you see is what you may read
The chat acts as you. It reads the organization you may read, and nodes that stand for a factory or workspace row are held to that row's own permission, row by row, exactly as on the System Map pages. A hidden node is not listed, followed or counted in a reason; the tools report how many rows they dropped, and the model is told to say so. Knowledge passages come only from sources whose scope you may open. Two people can ask the same question and get different answers, because they may read different things.
Like the concierge, it runs on the organization's model through SSO. When the token allowance is spent, or no model is available, SSO's own sentence is shown as the answer.
What it is not
- Not a change surface. It cannot create a card, move one, edit a node, start a rebuild or change a setting. Those are done on the pages, by a person, under the same permission checks as always.
- Not a live reading of the repository. It reads the map — the last analysed version — and the knowledge index, not the host. A commit that has not been analysed yet is not in its answers.
- Not a worker. What a station's agent reads when it works a card is the context pack, assembled by rule and recorded on the run. Ask your system is for people.
- Not the discussion. Conversations are yours alone; a decision reached in the chat is private to you until you put it on the card.