You’re not brief on instruments. Or fashions. Or frameworks.

What you’re brief on is a principled approach to make use of them — at scale.

Constructing efficient generative AI workflows, particularly agentic ones, means navigating a combinatorial explosion of selections.

Each new retriever, immediate technique, textual content splitter, embedding mannequin, or synthesizing LLM multiplies the area of attainable workflows, leading to a search area with over 10²³ attainable configurations. 

Trial-and-error doesn’t scale. And model-level benchmarks don’t replicate how parts behave when stitched into full programs.

That’s why we constructed syftr — an open supply framework for routinely figuring out Pareto-optimal workflows throughout accuracy, price, and latency constraints.

The complexity behind generative AI workflows

As an instance how rapidly complexity compounds, think about even a comparatively easy RAG pipeline just like the one proven in Determine 1.

Every part—retriever, immediate technique, embedding mannequin, textual content splitter, synthesizing LLM—requires cautious choice and tuning. And past these choices, there’s an increasing panorama of end-to-end workflow methods, from single-agent workflows like ReAct and LATS to multi-agent workflows like CaptainAgent and Magentic-One. 

What’s lacking is a scalable, principled solution to discover this configuration area.

That’s the place syftr is available in.

Its open supply framework makes use of multi-objective Bayesian Optimization to effectively seek for Pareto-optimal RAG workflows, balancing price, accuracy, and latency throughout configurations that will be not possible to check manually.

Benchmarking Pareto-optimal workflows with syftr

As soon as syftr is utilized to a workflow configuration area, it surfaces candidate pipelines that obtain sturdy tradeoffs throughout key efficiency metrics.

The instance beneath reveals syftr’s output on the CRAG (Complete RAG) Sports activities benchmark, highlighting workflows that keep excessive accuracy whereas considerably lowering price.

Whereas Determine 2 reveals what syftr can ship, it’s equally vital to grasp how these outcomes are achieved. 

On the core of syftr is a multi-objective search course of designed to effectively navigate huge workflow configuration areas. The framework prioritizes each efficiency and computational effectivity – important necessities for real-world experimentation at scale.

Since evaluating each workflow on this area isn’t possible, we sometimes consider round 500 workflows per run.

To make this course of much more environment friendly, syftr features a novel early stopping mechanism — Pareto Pruner — which halts analysis of workflows which might be unlikely to enhance the Pareto frontier. This considerably reduces computational price and search time whereas preserving end result high quality. 

Why present benchmarks aren’t sufficient

Whereas mannequin benchmarks, like MMLU, LiveBench, Chatbot Area, and the Berkeley Perform-Calling Leaderboard, have superior our understanding of remoted mannequin capabilities, basis fashions not often function alone in real-world manufacturing environments.

As a substitute, they’re sometimes one part — albeit an important one — inside bigger, refined AI programs.

Measuring intrinsic mannequin efficiency is vital, nevertheless it leaves open vital system-level questions: 

• How do you assemble a workflow that meets task-specific objectives for accuracy, latency, and price?
• Which fashions do you have to use—and during which components of the pipeline?

syftr addresses this hole by enabling automated, multi-objective analysis throughout total workflows.

It captures nuanced tradeoffs that emerge solely when parts work together inside a broader pipeline, and systematically explores configuration areas which might be in any other case impractical to guage manually.

syftr is the primary open-source framework particularly designed to routinely establish Pareto-optimal generative AI workflows that steadiness a number of competing goals concurrently — not simply accuracy, however latency and price as properly.

It attracts inspiration from current analysis, together with:

• AutoRAG, which focuses solely on optimizing for accuracy
• Kapoor et al. ‘s work, AI Brokers That Matter, which emphasizes cost-controlled analysis to stop incentivizing overly pricey, leaderboard-focused brokers. This precept serves as considered one of our core analysis inspirations. 

Importantly, syftr can also be orthogonal to LLM-as-optimizer frameworks like Hint and TextGrad, and generic move optimizers like DSPy. Such frameworks could be mixed with syftr to additional optimize prompts in workflows. 

In early experiments, syftr first recognized Pareto-optimal workflows on the CRAG Sports activities benchmark.

We then utilized Hint to optimize prompts throughout all of these configurations — taking a two-stage strategy: multi-objective workflow search adopted by fine-grained immediate tuning.

The end result: notable accuracy enhancements, particularly in low-cost workflows that originally exhibited decrease accuracy (these clustered within the lower-left of the Pareto frontier). These features counsel that post-hoc immediate optimization can meaningfully increase efficiency, even in extremely cost-constrained settings.

This two-stage strategy — first multi-objective configuration search, then immediate refinement — highlights the advantages of mixing syftr with specialised downstream instruments, enabling modular and versatile workflow optimization methods.

Constructing and increasing syftr’s search area

Syftr cleanly separates the workflow search area from the underlying optimization algorithm. This modular design allows customers to simply lengthen or customise the area, including or eradicating flows, fashions, and parts by modifying configuration information.

The default implementation makes use of Multi-Goal Tree-of-Parzen-Estimators (MOTPE), however syftr helps swapping in different optimization methods.

Contributions of recent flows, modules, or algorithms are welcomed through pull request at github.com/datarobot/syftr.

Constructed on the shoulders of open supply

syftr builds on a lot of highly effective open supply libraries and frameworks:

• Ray for distributing and scaling search over massive clusters of CPUs and GPUs
• Ray Serve for autoscaling mannequin internet hosting
• Optuna for its versatile define-by-run interface (much like PyTorch’s keen execution) and help for state-of-the-art multi-objective optimization algorithms
• LlamaIndex for constructing refined agentic and non-agentic RAG workflows

• HuggingFace Datasets for quick, collaborative, and uniform dataset interface
• Hint for optimizing textual parts inside workflows, similar to prompts

syftr is framework-agnostic: workflows could be constructed utilizing any orchestration library or modeling stack. This flexibility permits customers to increase or adapt syftr to suit all kinds of tooling preferences.

Case examine: syftr on CRAG Sports activities

Benchmark setup

The CRAG benchmark dataset was launched by Meta for the KDD Cup 2024 and contains three duties:

• Job 1: Retrieval summarization
• Job 2: Data graph and internet retrieval
• Job 3: Finish-to-end RAG

syftr was evaluated on Job 3 (CRAG3), which incorporates 4,400 QA pairs spanning a variety of subjects. The official benchmark performs RAG over 50 webpages retrieved for every query. 

To extend problem, we mixed all webpages throughout all questions right into a single corpus, making a extra real looking, difficult retrieval setting.

Key observations and insights

Throughout datasets, syftr constantly surfaces significant optimization patterns:

• Non-agentic workflows dominate the Pareto frontier. They’re sooner and cheaper, main the optimizer to favor these configurations extra regularly than agentic ones.
• GPT-4o-mini regularly seems in Pareto-optimal flows, suggesting it gives a robust steadiness of high quality and price as a synthesizing LLM.
• Reasoning fashions like o3-mini carry out properly on quantitative duties (e.g., FinanceBench, InfiniteBench), doubtless attributable to their multi-hop reasoning capabilities.
• Pareto frontiers ultimately flatten after an preliminary rise, with diminishing returns in accuracy relative to steep price will increase, underscoring the necessity for instruments like syftr that assist pinpoint environment friendly working factors.

  We routinely discover that the workflow on the knee level of the Pareto frontier loses only a few proportion factors in accuracy in comparison with essentially the most correct setup — whereas being 10x cheaper.

  syftr makes it straightforward to seek out that candy spot.

Price of working syftr

In our experiments, we allotted a finances of ~500 workflow evaluations per activity. Though actual prices range based mostly on the dataset and search area complexity, we constantly recognized sturdy Pareto frontiers with a one-time search price of roughly $500 per use case.

We count on this price to lower as extra environment friendly search algorithms and area definitions are developed.

Importantly, this preliminary funding is minimal relative to the long-term features from deploying optimized workflows, whether or not by means of diminished compute utilization, improved accuracy, or higher consumer expertise in high-traffic programs.

For detailed outcomes throughout six benchmark duties, together with datasets past CRAG, consult with the full syftr paper. 

Getting began and contributing

To get began with syftr, clone or fork the repository on GitHub. Benchmark datasets can be found on HuggingFace, and syftr additionally helps user-defined datasets for customized experimentation.

The present search area contains:

• 9 proprietary LLMs
• 11 embedding fashions
• 4 common immediate methods
• 3 retrievers
• 4 textual content splitters (with parameter configurations)
• 4 agentic RAG flows and 1 non-agentic RAG move, every with related hierarchical hyperparameters

New parts, similar to fashions, flows, or search modules, could be added or modified through configuration information. Detailed walkthroughs can be found to help customization.

syftr is developed absolutely within the open. We welcome contributions through pull requests, function proposals, and benchmark reviews. We’re notably interested by concepts that advance the analysis route or enhance the framework’s extensibility.

What’s forward for syftr

syftr remains to be evolving, with a number of lively areas of analysis designed to increase its capabilities and sensible affect:

• Meta-learning
  At present, every search is carried out from scratch. We’re exploring meta-learning methods that leverage prior runs throughout related duties to speed up and information future searches.
• Multi-agent workflow analysis
  Whereas multi-agent programs are gaining traction, they introduce further complexity and price. We’re investigating how these workflows evaluate to single-agent and non-agentic pipelines, and when their tradeoffs are justified.
• Composability with immediate optimization frameworks
  syftr is complementary to instruments like DSPy, Hint, and TextGrad, which optimize textual parts inside workflows. We’re exploring methods to extra deeply combine these programs to collectively optimize construction and language.
• Extra agentic duties
  We began with question-answer duties, a vital manufacturing use case for brokers. Subsequent, we plan to quickly increase syftr’s activity repertoire to code technology, information evaluation, and interpretation. We additionally invite the group to counsel further duties for syftr to prioritize.

As these efforts progress, we goal to increase syftr’s worth as a analysis device, a benchmarking framework, and a sensible assistant for system-level generative AI design.

In the event you’re working on this area, we welcome your suggestions, concepts, and contributions.

Attempt the code, learn the analysis

To discover syftr additional, take a look at the GitHub repository or learn the total paper on ArXiv for particulars on methodology and outcomes.

Syftr has been accepted to seem on the Worldwide Convention on Automated Machine Studying (AutoML) in September, 2025 in New York Metropolis.

We look ahead to seeing what you construct and discovering what’s subsequent, collectively.