Overview

Send Braintrust alerts to Callgoose SQIBS. A Braintrust alert runs a query on your AI application logs and posts to your SQIBS API endpoint when it matches, for example when error rates or low scores rise. SQIBS opens an incident for each alert.

You need
A Braintrust project with permission to create alerts, and a SQIBS API token
Setup time
About 10 minutes
Result
Braintrust alerts open SQIBS incidents

How it works

Braintrust
Alert fires
An alert's filter matches new logs.
Braintrust
Sends a webhook
The alert posts to its webhook, as JSON.
SQIBS
API filter matches
The filter decides whether to open or resolve, and the priority.
SQIBS
Incident opens
The incident takes its title from the alert name.

Want SQIBS to fix known issues before anyone is paged? See Fix known issues automatically.

Step 1: Add the API filter in SQIBS

  1. In SQIBS, open Services, choose the service that should receive these incidents, and click Add Or Update API Integration.
  2. Add a filter with these values (there is no built-in Braintrust template yet, so enter them by hand):
Part Payload JSON key Key value contains What it does
Trigger "automation"."name" Opens an incident.
Title from "automation"."name" Becomes the incident title.
Description from "details" Becomes the incident description.
  1. Choose the Priority for these incidents and click Save.

Braintrust does not send a message when the problem clears, so close these incidents in SQIBS (or with an automation workflow) when the fix is in.

Not sure of a key? Send one test message first (step below), then build the filter from the real payload.

Find the right key with the JSON key creator
In the API Integration window, click the round {} button at the bottom right. Paste a payload that SQIBS received (copy it from the API log). SQIBS lists every key with its value, each with a copy button. Copy the key into Payload JSON key and the value into Key value contains. No need to read the JSON by hand.

Step 2: Create an alert in Braintrust

  1. In the Braintrust project, open Configuration > Alerts and click + Alert.
  2. Give it a name, choose the event type, and write the filter (BTQL) that should alert, for example logs with an error.
  3. In the webhook URL field, enter your SQIBS endpoint:
https://<your-endpoint>/sqibs-api/v1/process?from=braintrust&token=<your-token>

Replace <your-endpoint> and <your-token> with the values from your SQIBS account. See how to get the endpoint URL and API tokens.

  1. Click Test to send a test alert, then Create alert.

Step 3: Test it

The API log saves payloads only while API Tracking is on for the token. Turn it on under API Tokens before you test; it switches itself off after 48 hours.

  1. Click Test on the alert.
  2. In SQIBS, a new incident opens. The payload below shows the shape to expect. Field names can differ between versions, so send one test message first and copy the exact keys from the API log with the JSON key creator.
{
  "organization": {
    "name": "example"
  },
  "project": {
    "name": "support-bot"
  },
  "automation": {
    "name": "Error rate on support-bot",
    "event_type": "logs"
  },
  "details": {
    "count": 14,
    "related_logs_url": "https://www.braintrust.dev/app/example/p/support-bot/logs"
  }
}

No incident? Open the API log (with API Tracking on). If the payload is there, the filter did not match it: paste it into the JSON key creator and check the keys. If it is not there, check the URL and token in the Braintrust alert.

Fix known issues automatically

Many alerts have a known fix: restart a service, clear a full disk, run a health check. SQIBS can run that fix for you the moment the incident opens. This is event-driven auto remediation. For example, when Braintrust reports a burst of failed model calls, SQIBS can run a workflow that switches the application to a fallback model.

Incident opens
Workflow starts
SQIBS runs the automation workflow linked to the filter.
Fixed
Incident resolved
The issue is fixed before anyone has to step in.
Not fixed
Escalated to the engineer
After the waiting time you set, the escalation policy pages on-call as usual.

Safe by design. SQIBS runs only workflows your own engineers wrote and approved for your environment. Nothing is guessed or generated: the same steps an engineer would run by hand, run sooner. If the workflow does not fix the issue, the engineer is paged as usual.

To set it up, in the same API filter:

  1. Select Workflow: choose the automation workflow to run when this filter opens an incident.
  2. If success then make incident resolved: turn on to resolve the incident when the workflow succeeds.
  3. Waiting time for escalate to escalation policies: how many minutes to give the workflow before paging the on-call engineer.

Start with low-risk workflows, such as a health check, and add fixes as you come to trust them.

Learn more: