Regex Match - Phoenix

Compiles a regex and reports a binary match / mismatch against the model’s output — handy for asserting format invariants like phone numbers, currency strings, UUIDs, or identifier prefixes without paying for a model call.

For a zero-code option, see the pre-built Regex Match metric. Use the version below when you want to extend the logic — capture groups, multiple alternative patterns, richer explanations.

Code#

import redef evaluate(output, pattern):    if output is None:        return {            "label": "missing",            "score": 0.0,            "explanation": "Output is missing.",        }    try:        compiled = re.compile(pattern)    except re.error as exc:        return {            "label": "invalid_pattern",            "score": 0.0,            "explanation": f"Failed to compile regex {pattern!r}: {exc}",        }    match = compiled.search(str(output))    if match is None:        return {            "label": "mismatch",            "score": 0.0,            "explanation": f"Output does not match /{pattern}/.",        }    return {        "label": "match",        "score": 1.0,        "explanation": f"Matched substring: {match.group(0)!r}",    }

function evaluate({ output, pattern }: EvaluatorParams) {  if (output == null) {    return {      label: "missing",      score: 0,      explanation: "Output is missing.",    };  }  let regex: RegExp;  try {    regex = new RegExp(pattern as string);  } catch (err) {    return {      label: "invalid_pattern",      score: 0,      explanation: `Failed to compile regex ${JSON.stringify(pattern)}: ${        (err as Error).message      }`,    };  }  const match = String(output).match(regex);  if (match === null) {    return {      label: "mismatch",      score: 0,      explanation: `Output does not match /${pattern}/.`,    };  }  return {    label: "match",    score: 1,    explanation: `Matched substring: ${JSON.stringify(match[0])}`,  };}

Both re.search and String.match find a match anywhere in the string. If you need the entire output to match, switch to re.fullmatch in Python, or anchor the pattern with ^...$ in TypeScript.

Input mapping#

Parameter	Bind to
`output`	The model output, usually `output`.
`pattern`	Use a literal regex string (e.g. `^[A-Z]{3}-\d{4}$`) when every example shares the same pattern, or a dataset path (e.g. `metadata.expected_pattern`) when patterns vary per example.

Output configuration#

Categorical labels:

Label	Score
`match`	`1.0`
`mismatch`	`0.0`
`missing`	`0.0`
`invalid_pattern`	`0.0`

Optimization direction: maximize.

Runtime requirements#

Setting	Value
Sandbox	Any — works in the in-process WebAssembly (Python) or Deno (TypeScript) backends.
Dependencies	None — uses `re` / built-in `RegExp`.
Internet access	Not required.
Environment variables	None.

Warning

Untrusted regex patterns can be vulnerable to catastrophic backtracking (ReDoS). If pattern is sourced from per-example dataset values rather than a fixed literal, prefer patterns from a trusted column you control.