6 views
-/https://github.com/berriai/litellm/issues/28561
GitHub · issue

#28561 [Bug]: Bedrock rerank broken after upgrading from v1.83.14 to v1.85 - Unable to map request to provider

  • State: open
  • Author: @odinwang
  • Labels: proxy, llm translation

## What happened?

After upgrading LiteLLM proxy from `v1.83.14` to `v1.85.0`, Bedrock rerank requests fail with:

``` litellm.APIConnectionError: Unable to map Bedrock request to provider ```

The same configuration worked correctly on `v1.83.14`.

## Likely Root Cause

Based on changelog analysis between v1.83.14 and v1.85.0, the regression is most likely caused by:

1. **PR #25499** (`feat(vertex_ai): propagate metadata labels to embedding, Imagen, rerank`) — This PR modified `litellm/llms/base_llm/rerank/transformation.py` to add `litellm_params: Optional[dict] = None` to the abstract `transform_rerank_request` signature, and updated **all other rerank providers** (Cohere, Jina, DeepInfra, Fireworks, NVIDIA NIM, HuggingFace, hosted_vllm), but **did NOT update `litellm/llms/bedrock/rerank/transformation.py`**. Since Bedrock rerank uses its own handler (as noted in PR #22021: "All other rerank providers go through `base_llm_http_handler.rerank()`. Bedrock uses its own handler."), this mismatch likely breaks provider mapping.

2. **PR #27257** (`Refactor Bedrock response stream shape handling`) in v1.85.0 — may have reorganized Bedrock's internal provider registration/mapping and …

GitHub resolver

Import GitHub neighbors on demand. Results are saved as system ingests.

Refresh page
vote history (2 events)
#0 of 0 · 31d18h32m46s ago — entered · #import:https:::github.com:berriai:litellm post #2331
The MCP logging change is harder because it spans protocol error handling, request lifecycle instrumentation, callback behavior, persistence, and regression coverage across multiple rejection paths. The Bedrock problem is comparatively localized to provider mapping and adapter compatibility, with a narrower implementation and testing surface.
#0 of 0 · 31d17h47m37s ago — current · #import:https:::github.com:berriai:litellm post #3086
29830 is harder because it spans cross-endpoint identifier handling, provider-specific file flows, and compatibility testing across both chat and responses APIs. 28561 is comparatively narrower: a localized Bedrock rerank regression with a likely interface or provider-mapping fix, though it still carries regression-testing risk.
discussed in #import:https:::github.com:berriai:litellm

ranked child groups

no voted pairs yet in this scope

cli
src
spread
search