NeuralTrust | Platform for Agent Security.

Clients send the same OpenAI embeddings body they would send to OpenAI. Auth and the consumer slug are the same as chat.

curl -X POST "https://<gateway-host>/<consumer-slug>/v1/embeddings" \
  -H "X-AG-API-Key: <your-api-key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "text-embedding-3-small",
    "input": ["Hello from TrustGate"]
  }'

On a Private data plane, add X-AG-Gateway-Slug as in the chat snippets.

Providers

Registry	Upstream
OpenAI	`{base_url}/embeddings` (default `https://api.openai.com/v1/embeddings`)
Azure OpenAI	`{endpoint}/openai/deployments/{model}/embeddings?api-version=…`
Mistral	`https://api.mistral.ai/v1/embeddings` (chat host unchanged)
Custom / OpenAI-compatible	`{base_url}/embeddings`
Cohere	Same OpenAI body, adapted to Cohere `/v2/embed`
Vertex AI	Gemini `:embedContent` (one input) or `:batchEmbedContents` (many)
Amazon Bedrock	Titan embed only (`InvokeModel` with `{inputText}`). Model IDs that contain `titan-embed`, including `eu.` / `us.` prefixes.

Anthropic, Gemini (AI Studio), Bedrock Claude / Nova / Titan text, and other chat-only providers do not embed. They are left out of the candidate pool. Bedrock rejects a non–Titan-embed model with a 400. Vertex embeddings ignore the chat provider_options.action — they always call the embed endpoints.

Routing

The consumer’s registries are filtered to those that advertise embeddings.
Model resolution and load balancing run on that pool.
Pinning a registry that cannot embed is a 400, not a failover.
An empty capable pool is a 503.

Attach at least one embeddings-capable registry to the consumer. A mixed chat + embeddings pool is fine: chat still uses every chat-capable member; embeddings skip the rest.