Skip to main content
Sphinx uses context-based post-training to specialize agents for an organization’s work. It stores the learned context outside the underlying model, where you can read and revise it.

Learn from tasks, then apply the context

Training uses synthetic tasks grounded in your sources. Sphinx evaluates agent attempts and uses the evidence to refine a knowledge graph: a connected collection of pages that people can browse like a wiki. The goal is to learn context that helps with new questions, including business conventions and exceptions. To measure generalization, evaluate on tasks held out from training.

How this differs from fine-tuning

Fine-tuning changes a model’s parameters. Sphinx’s method instead updates external knowledge supplied as context. This lets you inspect and correct the learned material without changing the underlying model’s weights. The external knowledge can be used across different training and inference models. Evaluate each model and agent configuration on your own tasks before relying on it.

Prepare useful source material

Provide definitions, policies, methods, and examples that explain your work. Include version history and effective dates when practices change. Use the ingest prompt to identify authoritative sources and details to preserve. After an update, review the proposed changes. When sources disagree, resolve the controversy with the knowledge owner.

Evaluate on representative work

Use questions that require applying context, including exceptions and ambiguous terminology. Check answer quality alongside cost and latency. Results depend on the task, model, available sources, and agent configuration. Use Benchmarks (beta) for recurring question-and-criterion evaluations in Sphinx. Contact Sphinx to activate it. For published methodology and experiments, see Sphinx’s research article.