Skip to main content

For AI Teams

You are a member of an AI team and you have an AI system to test. This system is internal to your network: maybe you are still developing it and haven't released it yet, maybe it's meant for internal use only. Either way, you want to connect it to an external testing platform through spectral-bridge.

warning

The testing platform must run a compatible relay server for you to use spectral-bridge.

What you run​

You run two processes on a machine inside your network that can reach your AI system:

  • Relay client
  • Adapter

The relay client opens a single outbound WebSocket connection to the testing platform's relay server. You don't need to implement it. Use the Python package we provide.

The adapter is an HTTP server that receives requests from the relay client on POST /v1/chat/completions (the OpenAI chat completion contract) and forwards them to your AI system.

If your system already exposes an OpenAI-compatible endpoint, the built-in pass-through adapter does this out of the box. Otherwise, you write a custom adapter that translates between the /v1/chat/completions contract and whatever interface your system exposes.

The adapter as a security boundary​

The adapter is a security boundary, not just a translation layer. The testing platform can only reach the relay client and the adapter, both of which you run and control. No other device or service inside your network is reachable unless your adapter code explicitly calls it.