AI providers

An AI provider lets a job analyze report data, write text, and detect anomalies. Anaphora supports any provider that implements the OpenAI-compatible API specification.

Overview

With an AI provider, a job can:

• Summarize dashboard data and identify trends
• Flag unusual patterns or values
• Generate human-readable insights
• Add AI-written content to notifications

Configuration

Adding a provider

1. In the sidebar, click AI Providers
2. Click Create AI Provider
3. Configure the connection:

Field	Description	Example
Name	Identifier for this provider	Production GPT-5
Provider	Type of AI provider	OpenAI, DeepSeek, Custom (OpenAI API compatible)
Endpoint	OpenAI-compatible base URL (only for a custom provider)	https://api.openai.com/v1
API key	Authentication token	sk-... (stored encrypted)
Model	Model to use. Suggestions need the API key.	gpt-5.2

4. Click Test
5. Click Save

Free edition

The Free edition allows one AI provider per space.

Provider from the environment

The environment can own one provider: AI_PROVIDER, AI_MODEL and AI_API_KEY, plus AI_ENDPOINT for a custom (OpenAI-compatible) service and AI_NAME for the name in the list. Anaphora creates the provider at the first start, in the default space, and updates it at every start to match the environment. The token budgets you set in the UI do not change. See Installation.

Copying a provider to another space

To copy or clone an AI provider into another space, you need admin access to the space it comes from.

Provider inheritance

• Each Space can have its own AI provider configuration
• Jobs within a Space use that Space's configured provider
• Spaces without providers cannot use AI features

Using AI in jobs

Job capture integration

To add an AI action to a job:

1. Edit or create a job
2. Go to the Capture tab
3. Turn on the Advanced switch to use actions
4. Add an AI action
5. Configure:

Field	Description
Variable name	Name of the variable that stores the AI output
Provider	Select the AI provider to use
Prompt	Define the prompt with instructions
Output type	text, number, html
Answer length cap	Maximum tokens in the answer. Empty leaves it to the provider
Context	Everything so far or Only ticked rows
Send the snapshots as images	Send the snapshots to the model as images (only for providers that read images)

What an AI action sends

The Context section of an AI action decides which earlier snapshots and variables go to the model. Each snapshot goes as its accessibility tree (the text structure of the page). The header of the table counts the rows that are sent. The section has two modes:

Mode	What is sent
Everything so far	Every earlier snapshot and variable, minus the rows you untick. Anything you add to the job later is sent too. This is the default.
Only ticked rows	Only the rows you tick. Anything you add to the job later stays out until you tick it.

Switching modes does not change which rows are ticked.

Under the table, Send the snapshots as images also sends the ticked snapshots as pictures. Only a provider whose vendor reads images can take them: OpenAI does, DeepSeek and custom providers do not. For those, the box is disabled and its tooltip says why. Each image costs tokens on every run.

Answer length cap limits the reply, in tokens. A number needs a few tokens; a report section may need a few hundred.

Send less, pay less

On a dashboard with many panels, "Everything so far" sends every panel to answer a question about one of them. Use "Only ticked rows" to send only what the prompt needs.

Example: dashboard summary

Context: Dashboard snapshot showing system performance metrics
Type: Text
Prompt: "Summarize the key metrics from this dashboard,
        highlighting any values that exceed normal ranges."

Output: "System performance remains stable with 99.8% uptime.
        CPU utilization averaged 45% with a peak of 78% at 14:30 UTC.
        Memory usage is trending upward (+12% week-over-week) and
        may require attention if the trend continues."

Example: anomaly detection

Context: Error rate visualization captured from Kibana
Type: Number
Prompt: "Identify any anomalies in the error rate data. Respond with 0 if none found. Return 1 if anomalies are detected."

Output: 0

Use the output to add a condition in your job workflow for alerting.

A number in a condition

A Conditional block can test the variable of an AI action with the output type number. The answer of the model must be a bare number, or hold exactly one number, for example Severity: 7 or 7/10. Any other answer fails the run with the error The AI answer is not a number. This way, the job never takes a branch on an answer that nobody can read. Ask the model for one number only, and set a small Answer length cap.

The Kibana AI Triage template

The Kibana AI Triage template builds a job that asks an AI provider whether a human must look at the errors now. Pick it under Jobs → Create Job. It needs an AI provider in the space. It is an advanced template, so the Free edition does not have it. Every 15 minutes, the job:

1. Counts the HTTP 5xx errors of the last hour and of the hour before, in Kibana Discover.
2. Counts all the requests of the last hour.
3. Takes a snapshot of the overview dashboard.
4. Asks the AI provider for a severity from 0 to 10. The prompt gives the rules that an on-call engineer applies: more errors with more traffic is load, not a failure, and a source that sends nothing is a 7.
5. Stops when the severity is below 7. Nobody is notified.
6. At 7 and above, asks the AI provider for a short briefing, which goes in the report with the counts and the dashboard.

The job sends the report at most once every 3 hours. When the space has exactly one AI provider, the new job uses it. With more than one, select the Provider in each AI action. With none, add an AI provider to the space first. The template uses the public Kibana demo: change the URLs to your own Kibana.

Token budgets

Every AI provider shows the tokens it spent in the last 24 hours and in the last 7 days (the Tokens 24h and Tokens 7d columns). You can give each window a budget. The windows roll: "the last 24 hours" is not "since midnight".

Usage counts every call: scheduled runs, manual runs, retries, tests, previews and the Test button on the provider form. Some OpenAI-compatible services answer without usage data. Those calls count as zero, and the tooltip of the column says that the total is a lower bound.

Setting a budget

1. Open the AI provider.
2. In the Token budgets card, tick Daily, Weekly, or both.
3. Enter the budget in thousands of tokens (k tokens).
4. Save.

When you tick a window, Anaphora proposes double your current spend in that window, and at least 500k tokens for a day or 2,000k tokens for a week. This way, a new budget does not pause everything on the next call. The hint under each field shows the spend and its share of the budget.

A weekly budget below the daily one is refused, because it can never apply. You can save a budget that is already used up. The card then says what will happen.

When a budget is reached

• The job that made the call is paused.
• Every further AI call on that provider is refused until the window rolls over or you raise the budget.
• A job that runs into a budget that another job used up is paused the same way.
• The token columns turn green, yellow and red as the budget fills. A provider that reached its budget shows Budget reached.
• On the Jobs list, a paused job shows that an AI budget paused it, so you can tell it apart from a job that someone switched off.

Resuming paused jobs

Raising the budget does not switch the paused jobs back on. On the AI Providers list, the Paused jobs column counts the jobs that the budget paused. Click Resume to switch those jobs back on, and only those.

Budget alerts

Tick Notify the operator when a budget is reached to send an alert over the channel of the health alerts. Tick Also warn early, at to send a warning at a percentage of the budget too (80% by default). Each pause of a job sends an alert that names the job. The early warning, and a "reached" alert with no job to pause (a test or a preview), go out at most once per window.

Choose the channel under Settings > Application > Health Monitoring. You can set it without switching health monitoring on. Without a channel, jobs are still paused, but nothing is sent.

Next steps

• Spaces: configure Space-level AI providers
• Composer: add AI blocks to reports
• Self-monitoring: monitor AI provider health