Skip to main content
RoutMe

Observability: see why an answer went wrong

Every message to RoutMe's agent is a turn with a short reference such as vt_8f3k2a. When something fails, the Health & issues view groups it into one issue with a plain-language explanation and a suggested fix. Alerts go to email, and signed notification webhooks reach your own tools. Message text is never recorded.

RoutMe chat widget in dark mode on a clinic site showing a request status card with four stages
What the visitor saw. Each turn behind it has a reference.

What you get with observability

  • Turn references

    A visitor who hits an error sees a reference they can quote. In the dashboard you look it up and read the steps in order: identity, model, tool, guard, render.

  • Health & issues

    Each API, database and MCP connection shows its status, and each widget shows open issues and its fallback rate. Repeated failures collapse into one issue with a title, summary and hint.

  • Alerts

    A new issue at or above your minimum severity, an escalation, or a spike triggers an alert, at most one per issue per hour. You can also turn on a daily digest.

  • Notification webhooks

    Send signed JSON to your own https endpoint with retries. Events are issue.new, issue.escalated, issue.spike, digest and test, signed with HMAC-SHA256.

How it works

  1. 1

    A turn starts

    The widget stream emits the turn reference right after the run starts, so the visitor and your team can refer to the same turn.

  2. 2

    Steps are recorded in the background

    Names, timings, token counts, stop reasons, HTTP status codes and guard actions are recorded. Recording never slows the reply.

  3. 3

    Problems become issues

    A failing tool, a rejected credential, a blocked origin or an ungrounded answer is deduplicated into one issue, with a trend and up to five sample turns.

  4. 4

    You get told, and you fix it

    The hint says what to do, for example replace the key and press Test. Resolved issues reopen if they come back.

Examples

The same capability across industries

Clinic
The scheduling database rejects its password. One issue appears, with a hint to update the credential, instead of a hundred failed chats.
Travel
The booking API slows down. The connection turns degraded, then failing after three failures in a row, and an alert goes out.
Real estate
An ungrounded answer is caught. The issue quotes the unsupported claim so you can add it as knowledge.
SaaS
A new marketing site is not on the allowed origins list. The issue says exactly which origin to add.
Banking
The webhook feeds your incident channel, so the on-call engineer sees issue.spike with a link to open the issue.
Retail
A customer quotes a reference after a failed answer, and support opens that exact turn.

Limits worth knowing

  • Turns and hourly counts are kept for 30 days. Issue summaries are kept.
  • Never recorded: message text, reply text, tool arguments and results, system prompts, identity tokens and API keys.
  • Alerts depend on you configuring email addresses or a webhook.

Documentation

Frequently asked questions

What is a turn reference?

A short identifier like vt_8f3k2a for one visitor message and the agent's handling of it. It appears at the end of an error message and in the inbox.

Does RoutMe store what visitors typed in the traces?

No. Turn traces keep step names, timings and error classes, not message text, replies or tool results.

How are alerts limited?

A job checks every minute, and at most one alert per issue per hour is sent. A muted issue never alerts.

How do I verify a notification webhook?

Compute the HMAC-SHA256 of the raw body with your webhook secret and compare it to the signature header. Failed deliveries retry up to three times.

What kinds of issue exist?

Model, tool, connector, guard, widget, stream, configuration, rate limit and authentication issues, each with a severity and a plain-language hint.

See it on your own data

Tell us what your customers ask. We will show the agent answering from your systems.