Skip to main content
Deep Search is the analyst loop for training a new signal. You describe a failure mode in plain English, Bento surfaces candidate runs that look like it, you label them match or no-match, and you turn those labels into a live classifier. It’s how a one-line description becomes a signal that detects the pattern from then on. A session is one training run. You open it with a natural-language query that names the failure mode you’re hunting. scope decides which side of the run to match on: both (the default), user, or assistant. Bento ranks your trajectories by how similar each is to that description, and you work down the ranked list, giving each candidate a label of match, no_match, or skip. Training reads the match labels; those positive examples are what the signal learns to detect.

​
Commands

​
Train a signal

Open a session. query is required; scope defaults to both. The response carries the id you pass to every command below:
Pull the candidates, ranked by similarity to the query:
Label one as match, no_match, or skip. Re-labeling the same trajectory overwrites your prior verdict:
Repeat for each candidate. Labeled runs drop out, so re-list for a fresh page. When you’ve collected enough match examples, train the signal. name is required, prompt optional. The response returns the new signal_id:
Training needs at least one match label to learn from. A session with no positive examples returns a 400 asking you to label more.
Check where the session stands at any point:
The new detector lives under signals from here. The CLI trains the detector; binding it into a firing alert rule is dashboard-only.

​
See also

Signals

Where a trained detector lives. Read it, watch its events, tune it.

Trajectories

The analyzed runs Deep Search ranks and you label.

Issues

The tracked problems a signal helps surface and group.