Docs Navigation

9. Cheat sheet

All ABS actions and evaluators on one page.

Actions

CategoryActionsWhen to use
Communicationsays, asks, informs, greets, clarifies, confirms, rejects, suggests, showsBot or user speaks
Executioncalls, submits, retrieves, stores, updatesBot invokes tools
Interactionselects, uploads, approvesUser interacts with UI
Delegationhands_offBot transfers to a human

Evaluations

TypeLevelWhat it verifies
containsStepText includes substring
exact_matchStepExact text match
regexStepPattern match
f1 / bleu / rougeStepGraded match against a reference (ground_truth)
schemaStepValid JSON structure
tool_callStepCorrect tool call
llm_judgeStepQualitative criteria
GroundednessStepResponse supported by context
RelevanceStepResponse addresses the query
CoherenceStepLogical flow and consistency
FluencyStepNatural language quality
sequenceChainStep order
eventuallyChainSomething happens at least once
neverChainSomething never happens
countChainSomething happens N times
withinChainA happens within N steps of B
variable_consistencyChainA value stays unchanged
HateUnfairnessStepNo hate speech or bias
ViolenceStepNo violence or threats
SexualStepNo explicit content
SelfHarmStepNo self-harm encouragement

Where the judgment runs: deterministic checks (contains, sequence, never, …) and the reference metrics (f1, bleu, rouge) run locally with no model. llm_judge runs on the built-in judge (OpenAI/Anthropic/Gemini), Azure, AWS Bedrock, Google Vertex AI, or AI Evaluator. Safety dimensions run on the built-in judge with a curated rubric — no criteria required. Pick the engine with --adapter.


← Back to tutorial index