SIARE Glossary

Definitions of key terms used throughout the SIARE documentation.

Core Concepts

SOP (Standard Operating Procedure)

In SIARE, an SOP is a complete pipeline configuration that defines how agents collaborate. Represented as ProcessConfig in code. An SOP includes roles, graph structure, prompts, and constraints.

Also called: Pipeline, ProcessConfig

Role

An individual agent within a pipeline. Each role has a model, tools, prompts, and defined inputs/outputs. In user-facing documentation, roles may be called “agents” for clarity.

Code: RoleConfig User-facing: Agent

Director

The AI service that analyzes pipeline performance and proposes improvements. Acts as the “brain” of the evolution system - both diagnosing weaknesses and suggesting mutations.

Code: DirectorService

Gene Pool

Storage for SOP versions and their performance history. Tracks ancestry (parent-child relationships), computes Pareto frontiers, and maintains diversity via QD Grid.

Code: GenePool

Mutation

A targeted change to an SOP designed to improve performance. SIARE supports 7 mutation types: PROMPT_CHANGE, PARAM_TWEAK, ADD_ROLE, REMOVE_ROLE, REWIRE_GRAPH, CROSSOVER, META_PROMPT_CHANGE.

Code: MutationType, MutationProposal

Evolution Terminology

Evolution Job

A complete evolution run with defined budget, constraints, and phases. Orchestrated by the EvolutionScheduler.

Generation

One iteration of the evolution loop: execute → evaluate → diagnose → mutate.

Pareto Frontier

The set of SOPs where no single SOP dominates all others across all metrics. Represents optimal trade-offs (e.g., high accuracy vs. low cost).

QD Grid (Quality-Diversity Grid)

A data structure that maintains diverse solutions across behavioral dimensions. Prevents evolution from converging to a single local optimum.

Fitness

The overall performance score of an SOP, typically a weighted combination of multiple metrics.

Metrics & Evaluation

EvaluationVector

A collection of metric scores for a single SOP execution. Contains individual metric values plus an aggregate score.

Code: EvaluationVector

LLM Judge

A metric that uses an LLM to evaluate quality. Common for subjective measures like “answer quality” or “relevance.”

Code: MetricType.LLM_JUDGE

Programmatic Metric

A metric computed by code (not LLM). Examples: latency, cost, word count.

Code: MetricType.PROGRAMMATIC

Runtime Metric

A metric that measures execution characteristics like latency or API cost.

Code: MetricType.RUNTIME

Pipeline Structure

Graph

The DAG (Directed Acyclic Graph) that defines execution order and data flow between roles.

Code: GraphEdge

Conditional Edge

A graph edge that only executes if a condition is met. Enables dynamic routing based on intermediate outputs.

Code: GraphEdge.condition

PromptGenome

The collection of prompts for all roles in an SOP. The “DNA” that gets mutated during evolution.

Code: PromptGenome

PromptConstraints

Rules that limit what evolution can change in a prompt. Protects critical instructions from modification.

Code: PromptConstraints

Tools & Adapters

Tool Adapter

A class that wraps external functionality (APIs, databases, search engines) for use by agents.

Base class: ToolAdapter

Tool Registry

A global registry of available tool adapters. Adapters register via the @register_adapter decorator.

Code: ToolRegistry

Prompt Evolution

TextGrad

A prompt optimization strategy that uses LLM-generated critiques as “gradients” to improve prompts.

Code: TextGradStrategy

EvoPrompt

A prompt optimization strategy based on evolutionary algorithms (genetic algorithms or differential evolution).

Code: EvoPromptStrategy

MetaPrompt

A prompt optimization strategy that uses LLM meta-analysis to identify patterns and propose improvements.

Code: MetaPromptStrategy

API & Infrastructure

Execution Trace

A complete record of one pipeline run, including all role inputs/outputs, timing, and costs.

Code: ExecutionTrace

ConfigStore

Persistent storage for SOPs, prompts, and configurations.

Code: ConfigStore

Circuit Breaker

Fault tolerance mechanism that prevents repeated failures from overwhelming the system.

Code: CircuitBreaker

Terminology Mapping

User-Facing Term	Code Term	Description
Agent	Role	An individual component in the pipeline
Pipeline	SOP / ProcessConfig	The complete configuration
Evolution	Mutation + Selection	The optimization process
Score	EvaluationVector	Performance metrics
Memory	Context	Information passed between agents