Configuration Reference

Complete reference for configuring SIARE. All settings can be specified via:

  1. YAML config file (config.yaml)
  2. Environment variables (take precedence)
  3. Python code (for programmatic configuration)

Quick Setup 

# Copy the example configuration
cp config.example.yaml config.yaml

# Set required environment variables
export OPENAI_API_KEY="sk-your-key-here"

# Or for Anthropic
export ANTHROPIC_API_KEY="sk-ant-your-key-here"

Configuration File Structure 

# config.yaml
llm_provider:
  type: "openai"
  default_model: "gpt-4o-mini"

tool_adapters:
  vector_search:
    backend: "chroma"

evolution:
  default_budget:
    maxEvaluations: 1000

storage:
  base_path: "./data"

logging:
  level: "INFO"

LLM Provider

OpenAI 

llm_provider:
  type: "openai"
  api_key: "${OPENAI_API_KEY}"  # Can use env var
  default_model: "gpt-4o-mini"
  judge_model: "gpt-4o"         # For evaluation
  architect_model: "gpt-4o"     # For mutations
Setting Type Default Description
type string required "openai", "anthropic", or "ollama"
api_key string env var API key (prefer env var)
default_model string "gpt-4o-mini" Default model for agents
judge_model string default_model Model for LLM Judge metrics
architect_model string default_model Model for Director mutations

Anthropic 

llm_provider:
  type: "anthropic"
  api_key: "${ANTHROPIC_API_KEY}"
  default_model: "claude-3-haiku-20240307"

Ollama (Local) 

llm_provider:
  type: "ollama"
  base_url: "http://localhost:11434"
  default_model: "llama3.2"

Environment Variables

Variable Description
OPENAI_API_KEY OpenAI API key
ANTHROPIC_API_KEY Anthropic API key
OLLAMA_BASE_URL Ollama server URL (default: http://localhost:11434)

Tool Adapters 

tool_adapters:
  vector_search:
    backend: "chroma"
    embedding_model: "sentence-transformers/all-MiniLM-L6-v2"
    dimension: 384
    backend_config:
      persist_directory: "./data/chroma"
Setting Type Default Description
backend string "memory" "memory", "chroma", "pinecone", "qdrant"
embedding_model string "all-MiniLM-L6-v2" Sentence transformer model
dimension int 384 Embedding dimension
backend_config object {} Backend-specific settings

Backend-Specific Config

ChromaDB:

backend_config:
  persist_directory: "./data/chroma"
  collection_name: "siare_docs"

Pinecone:

backend_config:
  api_key: "${PINECONE_API_KEY}"
  environment: "us-east-1-aws"
  index_name: "siare"

Qdrant:

backend_config:
  url: "http://localhost:6333"
  collection_name: "siare"

SQL Database 

tool_adapters:
  sql:
    connection_string: "postgresql://user:pass@localhost:5432/db"
    dialect: "postgresql"
    read_only: true
    max_rows: 1000
Setting Type Default Description
connection_string string required Database connection URL
dialect string auto "postgresql", "mysql", "sqlite"
read_only bool true Prevent write operations
max_rows int 1000 Max rows per query 
tool_adapters:
  web_search:
    provider: "duckduckgo"
    max_results: 10
Setting Type Default Description
provider string "duckduckgo" "google", "bing", "duckduckgo", "serper", "brave"
max_results int 10 Maximum search results
api_key string - API key (for paid providers)

Evolution Settings

Budget Limits 

evolution:
  default_budget:
    maxEvaluations: 1000
    maxLLMCalls: 5000
    maxCost: 100.0
    maxWallTime: 3600
Setting Type Default Description
maxEvaluations int 1000 Maximum task evaluations
maxLLMCalls int 5000 Maximum LLM API calls
maxCost float 100.0 Maximum cost in USD
maxWallTime int 3600 Maximum runtime in seconds

QD Grid 

evolution:
  qd_grid:
    complexity_bins: 10
    embedding_dimensions: 2
    embedding_bins: 10
Setting Type Default Description
complexity_bins int 10 Bins for complexity dimension
embedding_dimensions int 2 Number of embedding dimensions
embedding_bins int 10 Bins per embedding dimension

Constraints 

evolution:
  default_constraints:
    maxRoles: 10
    maxEdges: 20
Setting Type Default Description
maxRoles int 10 Maximum agents in SOP
maxEdges int 20 Maximum graph edges

Prompt Evolution Strategies

TextGrad

Textual gradient descent with LLM-generated critiques.

prompt_evolution:
  textgrad:
    learning_rate: 0.1
    backprop_depth: 3
    gradient_aggregation: "mean"
    model: "gpt-4o"
    temperature: 0.7
Setting Type Default Description
learning_rate float 0.1 How much to change prompts (0.0-1.0)
backprop_depth int 3 Trace depth for gradient computation
gradient_aggregation string "mean" "mean", "max", "weighted"
model string "gpt-4" LLM for gradient computation
temperature float 0.7 Creativity in gradient generation

EvoPrompt

Evolutionary algorithms for prompt populations.

prompt_evolution:
  evoprompt:
    population_size: 10
    mutation_rate: 0.3
    crossover_rate: 0.7
    selection_method: "tournament"
    tournament_size: 3
Setting Type Default Description
population_size int 10 Population size
mutation_rate float 0.3 Mutation probability
crossover_rate float 0.7 Crossover probability
selection_method string "tournament" "tournament", "roulette"
tournament_size int 3 Tournament selection size

MetaPrompt

LLM meta-analysis for targeted improvements.

prompt_evolution:
  metaprompt:
    model: "gpt-4o"
    temperature: 0.8
    max_suggestions: 5
Setting Type Default Description
model string "gpt-4o" Model for meta-analysis
temperature float 0.8 Creativity level
max_suggestions int 5 Max improvement suggestions

Adaptive Strategy Selection 

prompt_evolution:
  adaptive:
    enabled: true
    strategy_weights:
      textgrad: 0.4
      evoprompt: 0.3
      metaprompt: 0.3
    auto_adjust: true
Setting Type Default Description
enabled bool true Enable adaptive selection
strategy_weights object equal Initial strategy weights
auto_adjust bool true Auto-adjust weights based on success

Storage 

storage:
  base_path: "./data"
  configs: "./data/configs"
  genes: "./data/genes"
  logs: "./data/logs"
Setting Type Default Description
base_path string "./data" Base directory for all storage
configs string "./data/configs" Config store path
genes string "./data/genes" Gene pool storage path
logs string "./data/logs" Log file directory

Logging 

logging:
  level: "INFO"
  format: "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
  file: "./data/logs/siare.log"
Setting Type Default Description
level string "INFO" "DEBUG", "INFO", "WARNING", "ERROR"
format string standard Log message format
file string - Log file path (optional)

API Server 

api:
  host: "0.0.0.0"
  port: 8000
  workers: 4
  cors_origins:
    - "http://localhost:3000"
  rate_limit:
    requests_per_minute: 60
Setting Type Default Description
host string "0.0.0.0" Server bind address
port int 8000 Server port
workers int 1 Number of workers
cors_origins list ["*"] Allowed CORS origins
rate_limit.requests_per_minute int 60 Rate limit per client

Circuit Breaker 

circuit_breaker:
  failure_threshold: 5
  recovery_timeout: 30
  half_open_requests: 3
Setting Type Default Description
failure_threshold int 5 Failures before circuit opens
recovery_timeout int 30 Seconds before half-open
half_open_requests int 3 Test requests in half-open

Retry Configuration 

retry:
  max_retries: 3
  base_delay: 1.0
  max_delay: 60.0
  exponential_base: 2.0
Setting Type Default Description
max_retries int 3 Maximum retry attempts
base_delay float 1.0 Initial delay in seconds
max_delay float 60.0 Maximum delay in seconds
exponential_base float 2.0 Exponential backoff base

Complete Example 

# config.yaml - Production configuration example

llm_provider:
  type: "openai"
  default_model: "gpt-4o-mini"
  judge_model: "gpt-4o"
  architect_model: "gpt-4o"

tool_adapters:
  vector_search:
    backend: "chroma"
    embedding_model: "sentence-transformers/all-MiniLM-L6-v2"
    dimension: 384
    backend_config:
      persist_directory: "./data/chroma"

  sql:
    connection_string: "${DATABASE_URL}"
    read_only: true
    max_rows: 1000

  web_search:
    provider: "serper"
    max_results: 10

evolution:
  qd_grid:
    complexity_bins: 10
    embedding_dimensions: 2
    embedding_bins: 10

  default_budget:
    maxEvaluations: 1000
    maxLLMCalls: 5000
    maxCost: 100.0
    maxWallTime: 3600

  default_constraints:
    maxRoles: 10
    maxEdges: 20

prompt_evolution:
  adaptive:
    enabled: true
  textgrad:
    learning_rate: 0.1
    model: "gpt-4o"
  evoprompt:
    population_size: 10
    mutation_rate: 0.3

storage:
  base_path: "./data"
  configs: "./data/configs"
  genes: "./data/genes"
  logs: "./data/logs"

logging:
  level: "INFO"
  file: "./data/logs/siare.log"

api:
  host: "0.0.0.0"
  port: 8000
  workers: 4
  cors_origins:
    - "https://yourdomain.com"
  rate_limit:
    requests_per_minute: 60

circuit_breaker:
  failure_threshold: 5
  recovery_timeout: 30

retry:
  max_retries: 3
  base_delay: 1.0

Python Configuration

For programmatic configuration:

from siare.core.models import (
    ProcessConfig,
    EvolutionJob,
    MetricConfig,
    RetryConfig,
    CircuitBreakerConfig,
)


# Create retry config
retry_config = RetryConfig(
    max_retries=3,
    base_delay=1.0,
    max_delay=60.0,
    exponential_base=2.0,
)

# Create circuit breaker config
circuit_breaker_config = CircuitBreakerConfig(
    failure_threshold=5,
    recovery_timeout=30,
    half_open_requests=3,
)

# Create evolution job with budget
job = EvolutionJob(
    id="my_evolution",
    baseSopIds=["my_sop"],
    taskSet=my_tasks,
    metricsToOptimize=[...],
    budget=BudgetUsage(
        maxEvaluations=1000,
        maxLLMCalls=5000,
        maxCostUSD=100.0,
    ),
    maxGenerations=20,
    populationSize=5,
)

Environment Variable Reference

Variable Description Required
OPENAI_API_KEY OpenAI API key Yes (if using OpenAI)
ANTHROPIC_API_KEY Anthropic API key Yes (if using Anthropic)
OLLAMA_BASE_URL Ollama server URL No (default: localhost:11434)
DATABASE_URL SQL database connection No
PINECONE_API_KEY Pinecone API key No (if using Pinecone)
SERPER_API_KEY Serper search API key No (if using Serper)
SIARE_CONFIG_PATH Custom config file path No (default: ./config.yaml)
SIARE_LOG_LEVEL Override log level No

See Also

Back to top

Copyright © 2024 Synapti AI. Distributed under the MIT License.