Intelligence

Run reports say how each response went and engagement events say what users did. Intelligence says what the conversations were about and how they ended, by having a model read them. It is a plan feature and it runs with the LLM provider you configure, so the conversations are only ever sent to your own provider account.

What it produces

• Topics. A bounded set of subjects, up to the ceiling you set, each with the threads that belong to it, its share of the range, its trend against the previous range, and the rates that make it interesting: no reply, loops, unanswered questions, depth, users, resolved.
• Tasks. The things users try to get done, with how consistently the assistant handles them. A task's page lists its threads, and from it you can create a skill for the assistant.
• Unanswered questions. Questions the assistant did not answer, grouped and counted.
• Signals. Model judged signals per thread: no_reply when the assistant never answered a turn, loop when the user repeated or rephrased. They join the behavioural signals from engagement events (stopped, regenerated, copied, edited) as Threads filters.
• The judged mix. Sentiment (positive, neutral, negative), resolution, and the languages of the range.
• Junk. Threads too short to judge, by the threshold you set, are marked junk and left unjudged.

Every figure links into Threads filtered to the matching conversations, so the page is a way of reading the threads, not a replacement for them.

How it runs

The worker runs once an hour. For each project with the feature on, it takes the threads that changed since the last pass, in batches, classifies them against the existing topics, and writes topics, tasks, questions and signals in one transaction, so a failed pass leaves nothing half written. The page covers the last 28 days by default and, like every page, any range you pick; a trend is shown only when the previous range was classified at least half as fully as the current one.

Run now in Settings › Intelligence queues a pass for the new threads; Re-analyse reclassifies the last n days, for instance after changing the instructions or the topic ceiling. The analysis runs ledger lists every pass with the threads scanned and classified, the junk count, tokens and cost.

Configuration

Settings › Intelligence holds the provider and model, the junk threshold, test user prefixes, the topic ceiling and the classifier instructions; see Settings. Instructions are the place to tell the classifier your domain vocabulary, for instance which product names are one topic.

Evaluators

Evaluators are the same worker judging with your own prompts. Each rule scores a share of new threads on a numeric, categorical or boolean scale, and the verdicts are listed on the Intelligence page and attached to runs as scores. See Evaluators.

Reading the figures

Intelligence holds what the model judged; what users did stays on Engagement. A thread can be judged resolved and still have been stopped by the user, and the two readings are shown side by side on the thread's page rather than merged.