Skip to main content
Workflows is Synth’s product surface for building reliable LLM systems. You bring a dataset of real inputs and what “good” looks like. Workflows learns an optimized prompt graph and gives you a stable way to serve it. Unlike our infrastructure APIs (GEPA, MIPRO, SFT, RL), Workflows is designed to feel like software you ship, not machinery you configure.

​
What you get

  • Dataset‑in → graph‑out: upload an ADASTaskSet and we train a prompt graph end‑to‑end.
  • Built‑in judging: rubric, contrastive, or gold‑examples scoring without writing judge code.
  • Live progress: poll status or stream events/metrics while training runs.
  • Downloadable artifacts: fetch the best prompt snapshot for local use.
  • Production inference: call /graph/completions to run the optimized graph on new JSON inputs.
Today Workflows starts with GEPA under the hood because it’s the fastest way to improve graphs from data. We’ll add additional improvement paths over time while keeping the same dataset and API surface.

​
How it works

  1. Define your workflow as data: tasks are your real inputs; gold outputs and/or rubrics define success.
  2. Train: POST /api/adas/jobs starts a Workflows training run.
  3. Monitor: GET /api/adas/jobs/{adas_id} or stream events.
  4. Use: download the best prompt or serve it via /api/adas/graph/completions.
Get started here: product/workflows/quickstart.

​
SDK + cookbooks

  • Python SDK: see sdk/jobs/workflows for the ADASJob API.
  • Examples: see cookbooks/workflows/overview for style matching and generative workflows.

​
Pricing

Workflows training and inference are usage‑based. See pricing/workflows.