Skip to content

YAML Configuration Guide

MiroFlow uses a configuration system for customizing AI agents, tools, and benchmarks.

Configuration Structure

Configuration Directory
config/
├── agent_*.yaml                      # Agent configurations (quickstart, web demo, etc.)
├── benchmark_*.yaml                  # Benchmark configurations
├── prompts/                          # Prompt classes (.py and .yaml)
├── llm/                              # LLM provider configurations
├── benchmark/                        # Benchmark dataset settings
└── tool/                             # Tool configurations

Quick Start

Run a Single Task 

bash scripts/test_single_task.sh \
  --config config/agent_quickstart.yaml \
  --task-question "What is the first country listed in the XLSX file that have names starting with Co?" \
  --file-path data/FSI-2023-DOWNLOAD.xlsx

Run Benchmarks 

# GAIA validation with MiroThinker (8 runs)
bash scripts/benchmark/mirothinker/gaia-validation-165_mirothinker_8runs.sh

# BrowseComp English with MiroThinker (3 runs)
bash scripts/benchmark/mirothinker/browsecomp-en_mirothinker_3runs.sh

# Or run a single benchmark run directly
uv run miroflow/benchmark/run_benchmark.py \
  --config-path config/benchmark_gaia-validation-165_mirothinker.yaml \
  benchmark.execution.max_concurrent=30 \
  output_dir="logs/gaia-validation-165/run_1"

Core Configuration

Basic Agent Setup

Basic Agent Configuration (agent_quickstart.yaml)
defaults:
  - benchmark: example_dataset
  - override hydra/job_logging: none
  - _self_

entrypoint: main_agent

main_agent:
  name: main_agent
  type: IterativeAgentWithToolAndRollback
  max_turns: 30

  llm:
    _base_: config/llm/base_openai.yaml
    provider_class: GPT5OpenAIClient
    model_name: gpt-5
    max_tokens: 128000

  prompt: config/prompts/prompt_main_agent_benchmark.yaml

  tools:
    - config/tool/tool-reading.yaml

  input_processor:
    - ${input-message-generator}
  output_processor:
    - ${output-summary}
    - ${output-boxed-extractor}

input-message-generator:
  type: InputMessageGenerator
output-summary:
  type: SummaryGenerator
output-boxed-extractor:
  type: RegexBoxedExtractor

output_dir: logs
data_dir: "${oc.env:DATA_DIR,data}"

Standard Benchmark Configuration

Benchmark Configuration (benchmark_gaia-validation-165_mirothinker.yaml)
defaults:
  - benchmark: gaia-validation-165
  - override hydra/job_logging: none
  - _self_

entrypoint: main_agent
main_agent:
  name: main_agent
  type: IterativeAgentWithToolAndRollback
  max_consecutive_rollbacks: 5
  max_turns: 200

  llm:
    _base_: config/llm/base_mirothinker.yaml

  prompt: config/prompts/prompt_main_agent_benchmark.yaml

  tools:
    - config/tool/tool-code-sandbox.yaml
    - config/tool/tool-serper-search.yaml
    - config/tool/tool-jina-scrape.yaml

  tool_blacklist:
    - server: "tool-serper-search"
      tool: "sogou_search"

  input_processor:
    - ${file-content-preprocessor}
    - ${input-message-generator}
  output_processor:
    - ${output-summary}
    - ${output-boxed-extractor}
    - ${output-exceed-max-turn-summary}

LLM Providers

Available Providers

  • Claude: ClaudeOpenRouterClient (via OpenRouter), ClaudeAnthropicClient (direct)
  • OpenAI: GPTOpenAIClient, GPT5OpenAIClient
  • OpenRouter (Generic): OpenRouterClient - access any model via OpenRouter
  • OpenAI-Compatible: OpenAIClient - generic client for OpenAI-compatible APIs
  • MiroThinker: MiroThinkerSGLangClient
  • DeepSeek: via OpenRouterClient or OpenAIClient

See LLM Clients Overview for details.

Available Tools

Tool Options

  • tool-reasoning / tool-reasoning-os: Enhanced reasoning capabilities
  • tool-searching: Web search, Wikipedia, Archive.org, and retrieval
  • tool-searching-serper / tool-serper-search: Lightweight Google search via Serper
  • tool-reading: Document processing
  • tool-code-sandbox: Python code execution in E2B sandbox
  • tool-image-video / tool-image-video-os: Visual content analysis
  • tool-audio / tool-audio-os: Audio processing
  • tool-jina-scrape: URL scraping with LLM-powered info extraction
  • tool-browsing: Web browsing
  • tool-markitdown: Document to markdown conversion

See Tool Overview for configurations.

Advanced Features

Input/Output Processors

Available Processors
# Input processors
input_processor:
  - ${file-content-preprocessor}     # Pre-process file content
  - ${input-hint-generator}          # Generate hints using LLM
  - ${input-message-generator}       # Generate initial message

# Output processors
output_processor:
  - ${output-summary}                # Summarize conversation
  - ${output-boxed-extractor}        # Extract \boxed{} answers via regex
  - ${output-final-answer-extraction} # Extract final answer using LLM
  - ${output-exceed-max-turn-summary} # Summarize when max turns exceeded

Benchmark Settings

Benchmark Configuration (config/benchmark/)
name: "your-benchmark"
data:
  data_dir: "${data_dir}/your-data"
execution:
  max_tasks: null      # null = no limit
  max_concurrent: 3    # Parallel tasks
  pass_at_k: 1         # Attempts per task

Tool Blacklist

You can disable specific tools from a tool server:

Tool Blacklist
tool_blacklist:
  - server: "tool-serper-search"
    tool: "sogou_search"
  - server: "tool-code-sandbox"
    tool: "download_file_from_sandbox_to_local"

Environment Variables

Required .env Configuration
# LLM Providers
OPENROUTER_API_KEY="your_key"
ANTHROPIC_API_KEY="your_key"
OPENAI_API_KEY="your_key"
OAI_MIROTHINKER_API_KEY="your_key"
OAI_MIROTHINKER_BASE_URL="your_url"

# Tools
SERPER_API_KEY="your_key"
JINA_API_KEY="your_key"
E2B_API_KEY="your_key"

# Optional
DATA_DIR="data/"
CHINESE_CONTEXT="false"

Key Parameters

Parameter Description Default
temperature LLM creativity (0.0-1.0) 0.3
max_tokens Response length limit 32000
max_turns Conversation turns (-1 = unlimited) -1
max_consecutive_rollbacks Max consecutive rollbacks before stopping 5
max_concurrent Parallel benchmark tasks 3

Best Practices

Quick Tips

  • Start simple: Use agent_quickstart.yaml as a base
  • Tool selection: Choose tools based on your task requirements
  • API keys: Always use environment variables, never hardcode
  • Resource limits: Set max_concurrent and max_tokens appropriately
  • Benchmark configs: Use the benchmark_*_mirothinker.yaml configs for reproducing benchmark results

Documentation Info

Last Updated: February 2026 · Doc Contributor: Team @ MiroMind AI