Runtime Request Families

Verdictan gives public proxy, discovery, WebSocket, and MCP routes. Clients can keep their native request format with one gateway. This page compares the primary text, embedding, and audio proxy families. It does not contain all routes. Governance coverage is different for each route. Verify the policy path for each transport.

Primary text, embedding, and audio routes

Family	Route	Use it for	Response style
Chat completions	POST /v1/chat/completions	Use this family for OpenAI-style chat clients.	The response uses JSON or streaming.
Responses	POST /v1/responses	Use this family for new text integrations that use the Responses-style shape.	The response uses JSON or streaming.
Messages	POST /v1/messages	Use this family for Anthropic-style message envelopes and content blocks.	The response uses JSON or streaming.
Embeddings	POST /v1/embeddings	Use this family for vector generation.	The response uses JSON.
Audio transcription	POST /v1/audio/transcriptions	Use this family for speech-to-text.	The response uses JSON.
Audio speech	POST /v1/audio/speech	Use this family for text-to-speech.	The response uses binary audio.

Other live gateway surfaces

Surface	Route	Owning guidance
Legacy text completions	POST /v1/completions	This route gives legacy compatibility. Do not infer Chat Completions and Responses governance parity.
Moderation	POST /v1/moderations	Read Runtime moderation configuration.
Model discovery	GET /v1/models, GET /v1/models/:model_id	Read Runtime models configuration.
WebSocket proxy	GET /v1/chat/completions/ws, GET /v1/responses/ws	Read WebSocket Proxy.
Published-host MCP	GET /mcp, POST /mcp	Read Gateway MCP Surface.

These surfaces keep their owning authentication, publication, routing, and governance boundaries. These surfaces do not make the Chat Completions and Responses policy path universal.

Policy-stage limits by family

A request family cannot enforce all policy-chain stages. The gateway request-family registry lists the supported stages.

If the active chain has an unsupported stage, the gateway rejects the request before upstream dispatch with policy.incompatible_active_chain.

The gateway does not skip unsupported stages without notification.

Family	Route	Supported policy stages	Specified limitations
Chat completions	POST /v1/chat/completions	pre_request, post_request, pre_response	This family uses the full input and output policy path.
Responses	POST /v1/responses	pre_request, post_request, pre_response	This family uses the full input and output policy path.
Messages	POST /v1/messages	pre_request, post_request, pre_response	This family uses shared access, input, tool, output, and audit enforcement with an Anthropic-shaped body.
Embeddings	POST /v1/embeddings	pre_request	Does not support post_request or pre_response. Vector output has no output-policy path.
Audio transcription	POST /v1/audio/transcriptions	pre_request	This family does not support post_request or pre_response.
Audio speech	POST /v1/audio/speech	pre_request	Does not support post_request or pre_response. Output policy does not govern binary audio.
Legacy completions	POST /v1/completions	pre_request	This family does not support post_request or pre_response. Do not infer Chat Completions and Responses parity.
Moderation	POST /v1/moderations	pre_request	This family does not support post_request or pre_response.

How to select between text families

• If you want the Responses-style request and output model, use /v1/responses for new text-family work.
• If the client uses the OpenAI chat-completions shape, use /v1/chat/completions.
• If the caller must use the Anthropic Messages shape, use /v1/messages.

Chat Completions, Responses, and Messages use shared policy stages. Route shape, streaming translation, structured-output support, and provider support remain different. Test the selected family with the active chain.

WebSocket routes apply request and response policy to complete text frames. They reject binary frames and malformed JSON-like text. Read Streaming with SSE and WebSocket Gateway.

Structured outputs

On the OpenAI-style Chat Completions and Responses routes, structured outputs use response_format in the request body. The Anthropic-style Messages route does not advertise json_object or json_schema response formats. Do not add this OpenAI field to a Messages request.

{
  "response_format": {
    "type": "json_object"
  }
}

For a named schema, use this request:

{
  "response_format": {
    "type": "json_schema",
    "json_schema": {
      "name": "order_summary",
      "strict": true,
      "schema": {
        "type": "object"
      }
    }
  }
}

Request-time cache and session hints

Chat Completions and Responses support request-time cache and session hints. Do not infer equivalent handling on Messages from the shared field names.

Field	Location	What it does
cache_control	Request body	This field requests prompt-caching behavior when runtime policy allows it.
session_id	Request body	This field groups related requests onto the same session lane.
x-session-id	Request header	This field is the header form of the same session identifier.

If the request contains session_id and x-session-id, their values must match.

What you configure elsewhere

You do not select some runtime behavior for each request. Use runtime configuration docs for:

• Configure default provider and fallback behavior.
• Configure privacy and retention settings.
• Configure organization-level plugin defaults.
• Configure evaluation and traffic-mirroring controls.

Read Runtime Configuration for those defaults.

Next steps

• Read OpenAI Integration.
• Read Runtime Configuration.