The Eval family contains 4 operators. Grade outputs against assertions, regexes, and judge prompts.
Evaluates a CEL boolean expression against merged inputs and fails the step if it returns false.
Version 1.0.0 · Semantics passthrough · Idempotent
| Field | Type | Required | Default | Description |
|---|
expression | string | yes | — | CEL expression returning a boolean. Inputs are bound under ctx.inputs.* and inputs.*. |
target | string | no | — | Optional target value exposed as ctx.inputs.target. |
| Port | Type | Required | Description |
|---|
input | any | no | Input |
| Port | Type | Required | Description |
|---|
output | object | no | Output |
Scores a response against a rubric using an LLM judge. Returns a score, justification, and per-criterion results.
Version 1.0.0 · Semantics passthrough · Secrets OPENROUTER_API_KEY
| Field | Type | Required | Default | Description |
|---|
max_score | number | no | 10 | |
model | string | no | — | Model identifier (resolved against the customer catalog). |
response | string | no | — | Response to evaluate. Falls back to inputs.output if omitted. |
rubric | string | yes | — | Evaluation rubric or criteria. Embedded into the judge’s system prompt. |
| Port | Type | Required | Description |
|---|
input | string | no | Input |
| Port | Type | Required | Description |
|---|
output | object | no | Output |
Native regex match/find_all/replace against a target string.
Version 1.0.0 · Semantics passthrough · Idempotent
| Field | Type | Required | Default | Description |
|---|
mode | ”match” | “find_all” | “replace” | no | "match" | |
pattern | string | yes | — | Regular expression pattern (Rust regex syntax). |
replacement | string | no | — | Replacement string (required when mode=replace). |
target | string | no | — | String to apply the pattern to. Falls back to inputs.output if omitted. |
| Port | Type | Required | Description |
|---|
input | string | no | Input |
| Port | Type | Required | Description |
|---|
output | object | no | Output |
Scores output quality against a reference or rubric.
Version 1.0.0 · Semantics passthrough · Visibility advanced · Secrets OPENROUTER_API_KEY
| Field | Type | Required | Default | Description |
|---|
max_tokens | integer | no | 4096 | |
model | string | no | — | Model identifier (e.g. ‘deepseek/deepseek-r1’). |
reference | string | no | — | Optional reference answer. |
rubric | string | yes | — | Evaluation rubric or criteria. |
stop_sequences | string[] | no | — | |
temperature | number | no | 0.7 | |
top_p | number | no | — | |
| Port | Type | Required | Description |
|---|
input | string | yes | Input |
| Port | Type | Required | Description |
|---|
explanation | string | no | Explanation |
score | number | no | Score |
usage | object | no | Usage |
Every operator above is rendered from its JSON schema in /opt/glyph/schemas/operators/:
schemas/operators/eval.assertion.json
schemas/operators/eval.grade.json
schemas/operators/eval.regex.json
schemas/operators/eval.score.json