[bot] Cohere: Embed Jobs API (`client.embed_jobs.create()`) not instrumented

[braintrust-bot]

Summary

The Cohere Python SDK's Embed Jobs API — client.embed_jobs.create() / client.embed_jobs.get() / client.embed_jobs.list() (and the AsyncClient equivalents) — is not instrumented. This is Cohere's bulk/batch embedding execution surface: it launches an async job that reads a Dataset of type embed-input, runs the configured embedding model (model, input_type, embedding_types) over every record, and writes the resulting vectors to a new embed-output Dataset. It is the async, large-scale counterpart to the synchronous client.embed() call, which is instrumented.

Calls to client.embed_jobs.create() fall through uninstrumented today — the Cohere integration only patches BaseCohere/AsyncBaseCohere (chat, chat_stream, embed, rerank) and the v2 client and audio transcription client. No patcher touches the embed_jobs resource, so job creation produces zero Braintrust tracing (no span for the submitted job, its model/config, or its terminal status).

What is missing

Cohere method	Instrumented?
client.chat() / client.chat_stream() (v1 + v2)	Yes
client.embed() (v1 + v2)	Yes
client.rerank() (v1 + v2)	Yes
client.audio.transcriptions.create()	Yes
client.embed_jobs.create()	No
client.embed_jobs.get()	No
AsyncClient.embed_jobs.create()	No

At minimum, instrumentation should create a span for embed_jobs.create() capturing:

• Input: dataset ID, model, input_type, embedding_types, job name/truncate settings
• Output: job ID and initial status (processing/complete/failed)
• Metadata: model name, dataset reference

This mirrors the pattern already accepted for the analogous async-batch-execution gaps in sibling integrations: OpenAI Batch API (#295), Google GenAI Batch API (#332), and Mistral Batch Jobs API (#272) — this is the same class of gap (an async, job-based generative-execution API that bypasses the wrapper) applied to Cohere's bulk embeddings surface.

Braintrust docs status

not_found — The Cohere integration page documents only: chat completion spans (cohere.chat / cohere.chat_stream), tool call spans, embedding spans (cohere.embed), rerank spans (cohere.rerank), and audio transcription spans (cohere.audio.transcriptions.create). Embed Jobs, batch/bulk embeddings, the Classify API, and the v1 generate() endpoint are not mentioned anywhere on the page.

Upstream sources

• Cohere Embed Jobs API reference (POST /v1/embed-jobs): https://docs.cohere.com/reference/create-embed-job
• Cohere conceptual guide, "Batch Embedding Jobs with the Embed API": https://docs.cohere.com/v2/docs/embed-jobs-api
• Cohere Python SDK (cohere-python) exposes this as client.embed_jobs.create(...) / .get(...) / .list(...) on both cohere.Client and cohere.AsyncClient

Local files inspected

• py/src/braintrust/integrations/cohere/patchers.py — defines ChatPatcher, ChatStreamPatcher, EmbedPatcher, RerankPatcher (+ async/v2 variants) and TranscriptionsCreatePatcher/AsyncTranscriptionsCreatePatcher; no embed_jobs patcher of any kind
• py/src/braintrust/integrations/cohere/integration.py — CohereIntegration.patchers registers only the five patchers above; embed_jobs is absent from the registry
• py/src/braintrust/integrations/cohere/tracing.py — no embed_job wrapper functions exist
• py/src/braintrust/integrations/cohere/test_cohere.py — no test cases reference embed_jobs
• Repo-wide search for embed_job/embedjob (case-insensitive) returns zero matches outside this issue

Relationship to existing issues

Distinct from #488 (Cohere v1 generate()/generate_stream()) and #343 (Cohere Classify API) — both already open and unrelated to Embed Jobs. Same class of gap as the already-filed provider Batch API issues (#295, #332, #272), applied to a surface (Cohere Embed Jobs) none of those cover.