Business Rules

Business rules are policies you write once and Muster evaluates against every trace your agents produce. Each rule emits a Langfuse score (0 = failed, 1 = passed) so you can filter, alert, and dashboard on violations the same way you do any other score.

Four rule types

Type	What it checks	Example use
`INPUT_OUTPUT_CONSISTENCY`	Compare a field on the input to a field on the output	"The agent's response mentions the user_query topic"
`THRESHOLD_RANGE`	A numeric field is within bounds	"Latency under 5 seconds"
`PATTERN_FORMAT`	A field matches a regex or preset (EMAIL, URL, PHONE)	"Output never contains a credit-card number"
`COMPLETENESS`	Required fields are non-empty	"Every response includes a citations array"

Each rule lives in MusterBusinessRule with:

name and optional description (the score name auto-derives as BR: <name>)
checkType — one of the four above
ruleConfig — JSON config tailored to the type (which fields, comparators, presets, etc.)
targetTraceName — null to apply to every trace, or a specific trace name to scope it
enabled — flip to disable without deleting

How evaluation runs

trace completes
   └── businessRuleEvaluation worker enqueued
         └── fetch the trace + all enabled rules whose targetTraceName matches
               └── evaluate each rule against trace input/output/metadata
                     └── write a Langfuse score (0 or 1) with reason comment
                           — source: EVAL
                           — metadata: { source: "muster-business-rule" }

Scores are stored on the same table as every other Langfuse score, so they show up in trace detail view, score analytics, and dashboard widgets without any extra wiring.

Wire up your first rule

In Business Rules → New rule:

Name: Response stays on topic
Type: INPUT_OUTPUT_CONSISTENCY
Input field: user_query
Output field: response.text
Comparison: CONTAINS
Target trace name: leave empty to apply to every trace
Enabled: on

Save, and the next trace your agent emits will get scored as BR: Response stays on topic = 0 or 1. Filter the Traces table by score to triage failures.

Tips

Rules cost nothing to leave on — evaluation is cheap, scores are cheap to store. Lean toward enabling more rules and triaging via score filters rather than running fewer rules.
Use targetTraceName to keep rules scoped. A "must include citations" rule probably only applies to your research-agent trace, not your tool-call trace.
Disable instead of deleting when iterating — keeps history queryable and lets you turn the rule back on without reconfiguring it.

Business Rules

Four rule types

How evaluation runs

Wire up your first rule

Tips

On this page