Intelligence

Intelligence reads your conversations with the provider and model selected in Settings. It turns the result into topic, task, question, and conversation level signals that help you find what users ask for and where the assistant falls short.

The default Intelligence range is 28 days. Its dashboard views link the aggregate figures back to the matching conversations, so you can move from a trend to the threads behind it.

What Intelligence produces

The Intelligence page also shows Analysed, Junk, No reply, Loops, Unanswered, and Resolved tiles. On a thread's Details tab, Analysis shows its topic, task, phrasing, sentiment, resolution, language, turns, repeats, unanswered question, and whether it was classified.

Skills

For each recurring task, analysis drafts a skill with a name, a one sentence description of when to use it, and instructions. It reads the task's conversations and what went wrong in them, including comments end users left on a thumbs down and notes reviewers left on reviews marked Needs work.

On the task page, a person can edit the draft and press Publish. Update saves later edits, and Unpublish withdraws the skill. A newer draft never replaces published text. The page offers Load draft instead. Publishing does not require an LLM provider or an assistant. Create assistant remains a secondary action on the task page.

A backend reads published skills through GET /v1/projects/skills or the list_skills MCP tool.

How a pass runs

The worker passes every hour. It takes a queued request before its ordinary work, and a queued request starts within the hour. A normal pass advances from its keyset watermark through conversations that changed since the preceding pass. A reanalysis instead covers the requested historical day range.

Run now queues a normal pass. Re-analyse last 28 days queues a reanalysis for 28 days. A reanalysis accepts 1 to 90 days and defaults to 28 days. Only one unfinished request can exist for a project. A second request is refused with An analysis request is already waiting.

For each selected conversation, the worker applies the junk rule before it calls the model. It classifies the remaining conversations against the known topics, tasks, and questions. New entries beyond the topic ceiling fold into Other. The pass writes its classifications in one transaction, so it does not leave a partial result when the write fails.

Configure

Use an API key client to identify test traffic with a prefix configured under Test users.

import { AssistantCloud } from "assistant-cloud";

const cloud = new AssistantCloud({
  apiKey: process.env.ASSISTANT_API_KEY!,
  userId: "test_ci_42",
  workspaceId: "test_ci",
});

The test_ prefix entered under Test users makes every thread of this user junk. For a provider token, the user id is the token's sub.

The worker validates the stored configuration again before a pass; an invalid value falls back to the default above, and a project that never saved the page runs with every default.

The Instructions value is added to the classifier prompt in this form:

Instructions from the project owner:

What the pages show

Settings › Intelligence has a Runs ledger. It puts a pending request first as either Analysis run or Re-analysis of n days, with its queued or running state. It then shows recent passes with the day, threads scanned, junk count, model, tokens, and cost or status. All runs opens the analysis runs page.

The Intelligence page shows coverage for threads and analysed conversations, the last three runs, and cumulative run totals for runs, failures, tokens, and cost. The Analysis runs section holds the pass history, including the work scanned and classified, junk count, tokens, and cost for each pass.

A project without the feature sees Intelligence requires a paid plan and an upgrade path. Intelligence is included on Pro, Startup, and Enterprise plans. The demo project is synthetic and read only. Its Intelligence views do not call a model.

From your code

The read only MCP tool list_topics returns the topics in the authenticated project. It takes no input and orders the result by thread count, then by name. Each topic includes id, name, created_at, updated_at, and thread_count.

{ "topics": [] }

Costs and limits

Intelligence spends the key of the provider selected in its settings. The runs ledger records the tokens and cost of each pass, so the cost of a normal pass and a reanalysis can be read separately.

Troubleshooting