I am on the team. Here are some responses to your questions:

>Could I ignore all flow runs that have an estimated cost above a certain threshold, so that the overall cost of optimization is less? Suppose I choose an acceptable level of accuracy and then skip some costlier exploration stuff. Is there a risk it doesn't find certain optimal configurations even under my cost limit?

See the section on the Pareto Pruner in the paper where we evaluate the performance of ignoring runs on the optimization. It shows that ignoring these points does not lead to suboptimal configurations. We have at various times implemented fixed cost / latency thresholds at various times during development to reduce our cost exposure.

>How does the system deal with long context-length documents that some models can handle and others can't? Does this approach work for custom models?

That is part of why we jointly optimize the embedding and retriever settings. For models that only support shorter context lengths the retrieved documents will be truncated to avoid errors, if they leads to poor performance it should try and find better settings (e.g. by being more selective with a reranker).

>Suppose I want to create and optimize for my own LLM-as-a-judge metrics.

Please reach out on our github and we can discuss in code in more detail there. We are interested in expanding the options to make syftr more extensible but are curious exactly how people are interesting in using it for more methods, especially whether you are interested in exploring higher-dimensional spaces (accuracy + cost + additional metrics).

> Are you going to flesh out the docs? We plan to, we focused on writing the paper and explaining how syftr works. Now we plan to do more attention on the details of how to run experiments, customize etc.

> Any recommendations for creating the initial QA dataset for benchmarking? We used HuggingFace datasets as the dataset format. If there is interest on iterating on versions of the dataset on HuggingFace, that is something we could look at. You can see the current datasets here: https://huggingface.co/DataRobot-Research

We did two separate studies, one of accuracy and cost the other of accuracy and latency. Most of our studies used cost as the definition is pretty clear and less sensitive to environmental conditions (e.g. LLM provider changes) or definition (e.g. typical vs. worst case). But latency is a concern for many usecases, so we did want to investigate that problem as well.

Note we used Random LLM which had a 0.84 correlation with human labeling rather than 0.90 from only using gpt-4o-mini.