Summary

Calling cohere-embed-v4 (model: bedrock/eu.cohere.embed-v4:0) via the OpenAI-compatible /v1/embeddings endpoint with encoding_format: "float" (or any string value) results in a 400 Bad Request from Bedrock:

litellm.BadRequestError: BedrockException - {"message":"Malformed request. Please check your parameter/values and try again."}. Received Model Group=cohere-embed-v4

Root cause

LiteLLM maps the OpenAI param encoding_format: "float" (string) to Cohere's embedding_types field, but writes it as a string. Cohere's Bedrock embedding API requires embedding_types to be an array of strings (["float"]).

In litellm/llms/cohere/embed/transformation.py, around line 113:

for k, v in inference_params.items():
    transformed_request[k] = v  # type: ignore

The inference_params dict already contains embedding_types: "float" as a string at this point (from the upstream OpenAI-to-Cohere param mapping), and it's written through unchanged.

Reproduction

LiteLLM model config:

model_list:
  - model_name: cohere-embed-v4
    litellm_params:
      model: bedrock/eu.cohere.embed-v4:0
      aws_region_name: eu-central-1
    model_info:
      mode: embedding

Trigger:

curl -X POST http://localhost:4000/v1/embeddings \
  -H "Authorization: Bearer <key>" -H "Content-Type: application/json" \
  -d '{"model":"cohere-embed-v4","input":"test","encoding_format":"float"}'

Result: HTTP 400 with BedrockException - Malformed request.

Same request without encoding_format returns 200 OK, confirming the bug is specifically in this param translation.

The captured payload sent to Bedrock (verified via tracing):

{"input_type": "search_document", "texts": ["test"], "embedding_types": "float"}

Should be:

{"input_type": "search_document", "texts": ["test"], "embedding_types": ["float"]}

Real-world impact

This breaks all OpenAI SDK clients that pass encoding_format (default for many libraries — e.g., the mem0ai package hardcodes encoding_format: "float" in its OSS embedder). Any agent memory layer using Cohere embed-v4 via Bedrock through LiteLLM is broken.

Proposed fix

In litellm/llms/cohere/embed/transformation.py _transform_request:

for k, v in inference_params.items():
    if k == "embedding_types" and isinstance(v, str):
        v = [v]
    transformed_request[k] = v  # type: ignore

Alternatively (cleaner) — wrap during the OpenAI→Cohere param mapping wherever encoding_format becomes embedding_types.

Verified the one-line fix above resolves the issue locally — Cohere embed-v4 on Bedrock now returns 200 OK with the same client request.

Versions

• LiteLLM: latest in ghcr.io/berriai/litellm:main-stable (pulled 2026-04-17)
• Bedrock model: eu.cohere.embed-v4:0 (EU cross-region inference profile)
• Region: eu-central-1

Could someone please open a PR? The fix is small and isolated.