AI & Agent Configuration

Back to Settings Help
LLM Endpoint (OpenAI-compatible)

Configure any OpenAI-compatible endpoint. This is used by both the contextual helper widget and agent flow analysis. Leave blank to disable AI features.

Base URL of your OpenAI-compatible API. The path /chat/completions will be appended automatically.
Applied only when the selected model supports thinking/reasoning controls.
Stored locally. Never sent to SOBS servers.
Safety Guard Model

Every LLM request (helper and agent flows) is first passed through a guard model to block unsafe, off-topic, or prompt-injection content. Guard endpoint and model are required; when unavailable, requests are blocked (fail-closed).

Data Loss Prevention (DLP) Endpoint

Optional. When agent flows are about to create a GitHub issue, SOBS can first POST the issue text to a DLP endpoint to scan for PII or secrets. The DLP service must accept {"text": "…"} and return {"flagged": true|false}. Leave blank to skip DLP scanning.

GitHub Integration (Agent Flows)

Agent flows can optionally create or reuse a GitHub issue with sanitised details and, when enabled, formally assign the issue to GitHub Copilot cloud agent using GitHub's supported issue-assignment API. Requires a fine-grained personal access token with Metadata: Read plus Actions, Contents, Issues, and Pull requests: Read and write permission on the target repository. Dedupe runs before creation, and separate limits control issue creation versus Copilot assignment.

Use the Repositories wizard for guided setup and token validation checks.
Token expiry date not set
Used only by agent flows for issue creation. CVE/enrichment repo sources are managed in GitHub Repositories.
Hard ceiling for opening brand new GitHub issues after dedupe (max 20).
Relief valve for how many new Copilot starts SOBS may request in one hour.
Keep this low to avoid overlapping Copilot work on similar incidents.
Leave blank to let GitHub use the repository default branch for agent work.
Appended to the issue context when SOBS formally assigns an issue to Copilot.
Custom System Prompt (optional)

Override the default system prompt used for both the contextual helper and agent analysis calls. Leave blank to use the built-in SOBS observability expert prompt.

Model Pricing (USD per 1M tokens)

Used to estimate costs on the AI Transparency page. Built-in defaults are listed first, and models observed in your trace data are auto-added with inferred starting prices. Edit any row to override the pricing, or add a fully custom model. Prices are per 1 million tokens.

Rows marked inferred were auto-created from observed model names and seeded from the closest known pricing family until you adjust them.
Model $/1M In $/1M Out
How It Works
  • Contextual Helper Widget — A chat panel available on every page (click the button in the navigation). It is aware of your current page and recent observability context, and calls the configured LLM endpoint.
  • Guard Model — Before any LLM call, the input is first sent to the guard model and must return safe (or ALLOWED); otherwise the request is blocked.
  • Agent Flows — Configurable rules that automatically invoke the LLM when anomaly rules fire or tag rules match. Results appear in Agent Rules and can optionally create or reuse GitHub issues, then request Copilot work when the issue is eligible.
  • DLP — When enabled, issue text is scanned for PII/secrets before being posted to GitHub.
  • Dedupe & Work Control — Before opening new GitHub work, SOBS checks recent work items and open GitHub issues, asks the local LLM whether they are the same, related, or unrelated, and suppresses Copilot assignment when work is already in progress.
  • Tools/Thinking Models — AI helper tool actions and thinking controls are intended for instruct/tools-capable models.