Efficient Search for Pareto-optimal Flows

syftr is an agent optimizer that helps you find the best agentic workflows for a given budget. You bring your own dataset, compose the search space from models and components, and syftr finds the best combination of parameters for your budget. It uses advances in multi-objective Bayesian Optimization and a novel domain-specific "Pareto Pruner" to efficiently sample a search space of agentic and non-agentic flows to estimate a Pareto-frontier (optimal trade-off curve) between accuracy and objectives that compete like cost, latency, throughput.

Please read more details in our blogpost and full technical paper.

We are excited for what you will discover using syftr!

Libraries and frameworks used

syftr builds on a number of powerful open source projects:

• Ray for distributing and scaling search over large clusters of CPUs and GPUs

• Optuna for its flexible define-by-run interface (similar to PyTorch’s eager execution) and support for state-of-the-art multi-objective optimization algorithms

• LlamaIndex for building sophisticated agentic and non-agentic RAG workflows

• HuggingFace Datasets for fast, collaborative, and uniform dataset interface

• Trace for optimizing textual components within workflows, such as prompts

Installation

Please clone the syftr repo and run:

curl -LsSf https://astral.sh/uv/install.sh | sh
uv venv --python 3.12.7
source .venv/bin/activate
uv sync --extra dev
uv pip install -e .

or to use syftr as a library, install directly from PyPi:

pip install syftr

NOTE: syftr works as a library, but still needs easy access to config.yaml and study files you intend to run. Config file should be present as ~/.syftr/config.yaml, or in your current working directory. You can download sample config file to your ~/.syftr directory with this command

curl -L https://raw.githubusercontent.com/datarobot/syftr/main/config.yaml.sample \
     -o ~/.syftr/config.yaml

You also need studies to run syftr. You can write your own or download our example study with this command to current working directory

curl -L https://raw.githubusercontent.com/datarobot/syftr/main/studies/example-dr-docs.yaml > example-dr-docs.yaml

Required Credentials

syftr's examples require the following credentials:

• Azure OpenAI API key
• Azure OpenAI endpoint URL (api_url)
• PostgreSQL server dsn (if no dsn is provided, will use local SQLite)

To enter these credentials, copy config.yaml.sample to config.yaml and edit the required portions.

Additional Configuration Options

syftr uses many components including Ray for job scheduling and PostgreSQL for storing results. In this section we describe how to configure them to run syftr successfully.

• The main config file of syftr is config.yaml. You can specify paths, logging, database and Ray parameters and many others. For detailed instructions and examples, please refer to config.yaml.sample. You can rename this file to config.yaml and fill in all necessary details according to your infrastructure.
• You can also configure syftr with environment variables: export SYFTR_PATHS__ROOT_DIR=/foo/bar
• When the configuration is correct, you should be able to run examples/1-welcome.ipynb without any problems.
• syftr uses SQLite by default for Optuna storage. The database.dsn configuration field can be used to configure any Optuna-supported relational database storage. We recommend Postgres for distributed workloads.

Quickstart

First, run syftr check to validate your credentials and configuration. Note that most LLM connections are likely to fail if you have not provided configuration for them. Next, try the example Jupyter notebooks located in the examples directory. Or directly run a syftr study using the CLI syftr run studies/example-dr-docs.yaml --follow or with the API:

from syftr import api

s = api.Study.from_file("studies/example-dr-docs.yaml")
s.run()

Obtaining the results after the study is complete:

s.wait_for_completion()
print(s.pareto_flows)
[{'metrics': {'accuracy': 0.7, 'llm_cost_mean': 0.000258675},
  'params': {'response_synthesizer_llm': 'gpt-4o-mini',
   'rag_mode': 'no_rag',
   'template_name': 'default',
   'enforce_full_evaluation': True}},
   ...
]

LLM Configuration

syftr can be configured to use a wide variety of LLMs from a variety of LLM providers. These are configured using the generative_models section of config.yaml.

Each LLM provider has some different configuration options as well as some common ones. Let's look at an example using gpt-4.5-preview hosted in Azure OpenAI:

generative_models:
  # azure_openai Provider Example
  azure_gpt_45_preview:
    provider: azure_openai

    temperature: 0.0
    max_retries: 0

    # Provider-specific configurations
    deployment_name: "gpt-4.5-preview"
    api_version: "2024-12-01-preview"
    additional_kwargs:
      user: syftr

    # Cost example - options are the same for all models (required)
    cost:
      type: tokens                      # tokens, characters, or hourly
      input: 75
      output: 150.00
      # rate: 12.00

    # LLamaIndex LLMetadata Example - keys and defaults are the same for all models
    metadata:
      model_name: gpt-4.5-preview
      context_window: 100000
      num_output: 2048
      is_chat_model: true
      is_function_calling_model: true
      system_role: SYSTEM

Provider-specific options

All LLM configurations defined under generative_models: share a common set of options inherited from the base LLMConfig:

• cost: (Object, Required) Defines the cost structure for the LLM.
  • type: (String, Required) Type of cost calculation: tokens, characters, or hourly.
  • input: (Float, Required) Cost for input (e.g., per million tokens/characters).
  • output: (Float, Required if type is tokens or characters) Cost for output.
  • rate: (Float, Required if type is hourly) Average cost per hour.
• metadata: (Object, Required) Contains essential metadata about the LLM.
  • model_name: (String, Required) The specific model identifier (e.g., "gpt-4o-mini", "gemini-1.5-pro-001").
  • context_window: (Integer, Optional) The maximum context window size. Defaults to 3900.
  • num_output: (Integer, Optional) Default number of output tokens the model is expected to generate. Defaults to 256.
  • is_chat_model: (Boolean, Optional) Indicates if the model is a chat-based model. Defaults to false.
  • is_function_calling_model: (Boolean, Optional) Indicates if the model supports function calling. Defaults to false.
  • system_role: (String, Optional) The expected role name for system prompts (e.g., SYSTEM, USER). Defaults to SYSTEM.
• temperature: (Float, Optional) The sampling temperature for generation. Defaults to 0.0.

See LLM provider-specific configuration to configure each supported provider.

Embedding models

You may also enable additional embedding model endpoints:

local_models:
...
  embedding:
    - model_name: "BAAI/bge-small-en-v1.5"
      api_base: "http://vllmhost:8001/v1"
      api_key: "non-default-value"
      additional_kwargs:
        extra_body:
          truncate_prompt_tokens: 512
    - model_name: "thenlper/gte-large"
      api_base: "http://vllmhost:8001/v1"
      additional_kwargs:
        extra_body:
          truncate_prompt_tokens: 512

Models added in the config.yaml will be automatically added to the default search space, or you can enable them manually for specific flow components.

Custom Datasets

See detailed instructions here.

Citation

If you use this code in your research please cite the following publication.

@article{syftr2025,
  title={syftr: Pareto-Optimal Generative AI},
  author={Conway, Alexander and Dey, Debadeepta and Hackmann, Stefan and Hausknecht, Matthew and Schmidt, Michael and Steadman, Mark and Volynets, Nick},
  booktitle={Proceedings of the International Conference on Automated Machine Learning (AutoML)},
  year={2025},
}

Contributing

Please read our contributing guide for details on how to contribute to the project. We welcome contributions in the form of bug reports, feature requests, and pull requests.

Please note we have a code of conduct, please follow it in all your interactions with the project.