Graphs API, Graph Gen Workflows & Verifier Graphs

TL;DR

Unified Graphs API for running optimized graphs and verifiers
Graph Gen workflows for dataset-in, graph-out training
Verifier graphs for trace scoring with structured rewards
RLM graphs for massive-context inference
Live monitoring dashboard for graph execution
Expanded model provider support for graphs and judging
VLM judge support for multi-modal evaluation
Hosted SWE agent judges
Clearer onboarding for Graph Gen, GEPA, GSPO, SFT, and verifier training

Graphs API: One Inference Surface

A single production API now handles graph inference for optimized graphs and verifiers. This simplifies how users ship graphs and judge traces while keeping validation and monitoring consistent. Start with the Graphs overview and graph inference. The SDK entry point is in the public repo: synth-ai graphs completions client.

Graph Gen: Workflows From Datasets

Graph Gen is the dataset-in, graph-out product surface for building reliable LLM workflows. It supports policy graphs, verifier graphs, and RLM graphs without requiring users to manage task apps. Use the Workflows overview, quickstart, and judging guide. SDK references: Graph Gen job API and Graph Gen dataset schema.

Verifier Graphs

Verifier graphs provide calibrated scoring for traces, returning scores, reasoning, and structured rewards. Start with Judging in Graph Gen, the Graphs overview, and the V3 trace format. Cookbook references: Verifier evolution benchmarks and Crafter verifier training.

RLM Graphs (Massive Context)

RLM graphs handle large contexts by materializing content and searching locally instead of stuffing prompts. See RLM graphs and the public Graph Evolve README. Cookbook references: RLM cookbook and RLM evaluation examples.

Monitoring: Graph Execution

A live monitoring dashboard now provides real-time visibility into graph execution, latency, and failures. Start from the Workflows overview and Graphs overview.

Judges: VLM + SWE

We added VLM judge support for multi-modal evaluation and hosted SWE agent judges for software engineering tasks. See Judging in Graph Gen and the public verifier resources: Verifier evolution cookbook.

Provider Support

Expanded model provider coverage for graph execution and judging (see model support and prompt optimization models). Public SDK reference: synth-ai README.

UX: End-to-End Paid Product Flows

We tightened onboarding across paid products: explicit task app requirements, clearer first-run paths for Graph Gen, GEPA, GSPO, SFT, and consistent language for rollouts and inference. Cookbook reference: GEPA SDK examples.

Cookbooks