Skip to main content

Anthropic, OpenAI, And BYOM

Relay can use providers for:

  • Product document generation.
  • Document-readiness questions.
  • Technical discovery questions.
  • Task generation.
  • Embeddings.

Embeddings are used for Relay's semantic context index. When source material is ingested, Relay extracts text, splits it into chunks, and stores a vector embedding for each chunk in Postgres with pgvector. Those chunks are then available as grounded context for document generation, quality review, and later retrieval and ranking workflows.

In the current product flow, embeddings do not train a model, do not expose source material to Slack, Linear, or GitHub, and do not make implementation decisions by themselves. They are an indexing layer that helps Relay find and score relevant source context.

If an embedding provider is unavailable in local or test environments, Relay can still store deterministic development embeddings so the chunking path continues to work.

Default Provider Disclosure​

Default provider disclosure shown in Settings:

  • Generation: Anthropic Claude.
  • Discovery questions: Anthropic Claude when enabled, deterministic templates otherwise.
  • Embeddings: OpenAI text-embedding-3-small.

BYOM settings can route generation and embeddings through Azure OpenAI, AWS Bedrock, or a custom provider where configured. Cost caps can pause generation when spend reaches the configured limit.