You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Gateway proxy: serve /v1/embeddings for OpenAI and more providers #503
Serve embeddings through the gateway proxy so clients can call POST /{consumer}/v1/embeddings against more upstream providers (same governance/observability path as chat).
Out of scope: semantic cache / LB internal embedding creators (pkg/infra/embedding/*). Victor (2026-08-22): do not expand that stack for this issue.
Invoker: adapts via AdaptEmbeddingRequest/Response, then calls providers.EmbeddingsClient.Embeddings
(pkg/app/proxy/provider.go)
Adapters: OpenAI embeddings + Cohere embed (+ rerank on /v1/rerank)
Cohere only implements EmbeddingsClient today — functional test TestCohereProvider_EmbeddingsCrossFormat proves OpenAI-shaped client → Cohere upstream
Gap
Calling /v1/embeddings against OpenAI / Azure / openaicompat / other registries fails with provider does not support embeddings because those clients only implement chat (Completions / CompletionsStream).
Target format today: Cohere → FormatCohereEmbed; everyone else → FormatOpenAIEmbeddings.
Work (PR-only, after #502 chat providers if desired — but this can ship independently)
OpenAI client: implement Embeddings → POST …/v1/embeddings (honour base_url)
Azure: embeddings deployment URL
openaicompat:…/embeddings on configured base_url (Together/Fireworks/vLLM/Ollama/etc.)
changed the title [-]Support embeddings endpoint across more providers[/-][+]Gateway proxy: serve /v1/embeddings for OpenAI and more providers[/+]on Aug 22, 2026
Architecture alignment (2026-08-22, shared with #506–#510)
When implementing OpenAI/Azure/openaicompat embeddings, prefer pre-filtering the consumer’s CandidateSet to registries whose provider implements EmbeddingsClient, instead of only failing at invoke with a retryable “does not support embeddings” (today’s pattern). Pinned incapable registry should be terminal. Keeps embeddings consistent with the multimodal capability model: capability on provider → registry inherits → consumer pool filters.
Activity
[-]Support embeddings endpoint across more providers[/-][+]Gateway proxy: serve /v1/embeddings for OpenAI and more providers[/+]vgmartinez commentedon Aug 22, 2026
Architecture alignment (2026-08-22, shared with #506–#510)
When implementing OpenAI/Azure/openaicompat embeddings, prefer pre-filtering the consumer’s CandidateSet to registries whose provider implements
EmbeddingsClient, instead of only failing at invoke with a retryable “does not support embeddings” (today’s pattern). Pinned incapable registry should be terminal. Keeps embeddings consistent with the multimodal capability model: capability on provider → registry inherits → consumer pool filters.