/v1/batches accepts /v1/chat/completions only. /v1/embeddings, /v1/responses,
/v1/completions, /v1/moderations, /v1/images/* and /v1/videos are rejected at
parse time, where the upstream API accepts all eight.
Bedrock batch inference does cover embedding models, so at least /v1/embeddings looks
reachable today — embeddings are also the endpoint where batching pays best, since a
corpus is exactly the workload a customer wants off the synchronous path.
Feature: probe Bedrock batch inference for embedding jobs and for Responses-shaped jobs,
then widen the accepted endpoint set to whatever it actually serves. Keep rejecting the
rest with the existing clean error rather than accepting and failing later.
/v1/batchesaccepts/v1/chat/completionsonly./v1/embeddings,/v1/responses,/v1/completions,/v1/moderations,/v1/images/*and/v1/videosare rejected atparse time, where the upstream API accepts all eight.
Bedrock batch inference does cover embedding models, so at least
/v1/embeddingslooksreachable today — embeddings are also the endpoint where batching pays best, since a
corpus is exactly the workload a customer wants off the synchronous path.
Feature: probe Bedrock batch inference for embedding jobs and for Responses-shaped jobs,
then widen the accepted endpoint set to whatever it actually serves. Keep rejecting the
rest with the existing clean error rather than accepting and failing later.