Relevant issues

Fixes #24641
Supersedes #24646 (OpenAI-native approach with test coverage; see "Why a new PR" below)

Linear ticket

N/A — external contributor

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory — this PR adds 6 new tests under tests/llm_translation/ (the long-standing home for Bedrock/Anthropic translation tests). Happy to relocate to tests/test_litellm/ if maintainers prefer.
• My PR passes all unit tests on make test-unit — verified locally on translation-unit subsets (see CI links below once runs populate).
• My PR's scope is as isolated as possible, it only solves 1 specific problem — PDF (OpenAI file content) passthrough on the tool-result path for Bedrock Converse and direct Anthropic.
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

• Branch creation CI run
  Link:
• CI run for the last commit
  Link:
• Merge / cherry-pick CI run
  Links:

Screenshots / Proof of Fix

Before (current litellm_internal_staging / v1.83.7-stable)

When a caller sends a PDF as an OpenAI Chat Completions file content block inside a tool message:

{"role": "tool", "tool_call_id": "...", "content": [
    {"type": "file", "file": {"file_data": "data:application/pdf;base64,...", "filename": "x.pdf"}},
]}

…the translated Bedrock toolResult.content is an empty list [], and the Anthropic tool_result content has zero document blocks. The model sees empty content and responds as if the tool returned nothing.

After

The same input now translates to:

• Bedrock: toolResult.content[0].document with format: "pdf" and the base64 payload as source bytes.
• Anthropic: a {"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": ...}} block inside the tool_result content.

Test output

All 6 new tests + the existing test_bedrock_converse_translation_tool_message pass:

tests/llm_translation/test_bedrock_completion.py::test_bedrock_converse_translation_tool_message PASSED
tests/llm_translation/test_bedrock_completion.py::test_bedrock_tool_message_openai_file_pdf_becomes_document PASSED
tests/llm_translation/test_bedrock_completion.py::test_bedrock_tool_message_image_url_pdf_data_uri_becomes_document PASSED
tests/llm_translation/test_bedrock_completion.py::test_bedrock_tool_message_image_url_png_still_becomes_image PASSED
tests/llm_translation/test_anthropic_completion.py::test_anthropic_tool_result_openai_file_pdf_becomes_document PASSED
tests/llm_translation/test_anthropic_completion.py::test_anthropic_tool_result_image_url_pdf_data_uri_becomes_document PASSED
tests/llm_translation/test_anthropic_completion.py::test_anthropic_tool_result_image_url_png_still_becomes_image PASSED

Type

🐛 Bug Fix

Changes

Root cause

The OpenAI → Bedrock and OpenAI → Anthropic tool-result converters in litellm/litellm_core_utils/prompt_templates/factory.py only handled text and image_url content types on the tool-message path. Two concrete bugs:

1. Bedrock _convert_to_bedrock_tool_call_result: No branch for type: "file". Additionally, the image_url branch called BedrockImageProcessor.process_image_sync, which correctly returns a {"document": ...} block for data:application/pdf;base64,... URIs — but the wrapper only appended when "image" in _block, so document blocks were silently dropped.

2. Anthropic convert_to_anthropic_tool_result: No branch for type: "file". The image_url branch unconditionally called create_anthropic_image_param, wrapping any data URI (including application/pdf) as type: "image" — which Anthropic rejects for non-image mime types.

User-message paths already handle type: "file" correctly (via anthropic_process_openai_file_message for Anthropic and _process_file_message for Bedrock); the gap was specifically on tool-result content.

Fix

Bedrock (_convert_to_bedrock_tool_call_result):

• Added a type: "file" branch that delegates to BedrockImageProcessor.process_image_sync and wraps the returned block as document or image appropriately.
• Added a document passthrough in the existing image_url branch so processor-produced document blocks (PDFs sent as image_url data URIs) are no longer dropped.

This single fix covers both the sync (_bedrock_converse_messages_pt) and async (_bedrock_converse_messages_pt_async) paths because both funnel tool-result translation through this helper.

Anthropic (convert_to_anthropic_tool_result):

• Added a type: "file" branch that delegates to the existing anthropic_process_openai_file_message helper.
• In the image_url branch, branched on the data-URI mime type: for application/* and text/* URIs, routes through the same file helper to produce a proper document block; otherwise keeps existing image handling. Added new helper _is_anthropic_document_data_uri.
• Extended AnthropicMessagesToolResultParam.content union in litellm/types/llms/anthropic.py to accept AnthropicMessagesDocumentParam (previously text + image only), reflecting what the Anthropic API actually accepts inside tool_result content.

Why a new PR (superseding #24646)

#24646 has been stalled for ~a month (blocked, no reviews, no tests). It handles the same user-facing bug by adding a type: "document" branch — but that's not a documented OpenAI Chat Completions content type. It forces callers to:

• Speak provider-native document shapes in the OpenAI payload
• Branch on target provider (Bedrock's document sub-schema differs from Anthropic's)

This PR takes the OpenAI-native approach: callers send the documented OpenAI Chat Completions file content shape for PDFs, and LiteLLM handles the provider-specific translation. Also includes test coverage (the #1 reason #24646 stalled).

Happy to close this PR if maintainers prefer to revive #24646 instead.

Tests added

Six new tests under tests/llm_translation/, all driven via the existing public translation entrypoints (_bedrock_converse_messages_pt and convert_to_anthropic_tool_result):

Bedrock:

• test_bedrock_tool_message_openai_file_pdf_becomes_document — new behavior.
• test_bedrock_tool_message_image_url_pdf_data_uri_becomes_document — regression for the silent-drop bug.
• test_bedrock_tool_message_image_url_png_still_becomes_image — locks in existing image-block behavior.

Anthropic:

• test_anthropic_tool_result_openai_file_pdf_becomes_document
• test_anthropic_tool_result_image_url_pdf_data_uri_becomes_document
• test_anthropic_tool_result_image_url_png_still_becomes_image

Each new-behavior test was developed via strict red→green TDD — verified failing against the unpatched code first, then implemented minimal code to pass.