fix: map Chat Completion file type to Responses API input_file

[gambletan]

Summary

Fixes #23588

When bridging /chat/completions to the Responses API, content items with type: "file" were falling through to the default handler in _convert_content_to_responses_format and being stringified as input_text. This caused the model to receive the Python dict representation as plain text instead of the actual file bytes.

Root cause: In litellm/completion_extras/litellm_responses_transformation/transformation.py, the _convert_content_to_responses_format method handles "text", "image_url", and "image" content types explicitly, but "file" was not handled — it fell through to the else branch which calls str(item) and wraps it as input_text.

Fix: Add an explicit elif item_type == "file" branch that maps:

{"type": "file", "file": {"file_data": "data:application/pdf;base64,...", "filename": "test.pdf"}}

to:

{"type": "input_file", "file_data": "data:application/pdf;base64,...", "filename": "test.pdf"}

This is the inverse of the existing _transform_input_file_item_to_file_item in the responses-to-completion direction, which already correctly maps input_file → file.

Test plan

• Added test_convert_chat_completion_file_type_to_input_file — verifies file_data and filename are correctly extracted
• Added test_convert_chat_completion_file_type_with_file_id — verifies file_id variant works

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 14, 2026 1:20pm

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes a bug where Chat Completion messages containing {"type": "file", ...} content items were incorrectly stringified as input_text when bridging to the Responses API, instead of being properly converted to input_file objects. The fix adds an explicit elif item_type == "file" branch in _convert_content_to_responses_format that correctly flattens the nested file dict into a flat input_file structure expected by the Responses API.

Key changes:

• Added elif item_type == "file" handling in LiteLLMResponsesTransformationHandler._convert_content_to_responses_format (lines 699–712) that maps file_id, file_data, and filename from the nested file dict into a flat input_file object.
• Two new mock-only unit tests cover the file_data+filename and file_id variants of the file content item.
• Remaining changes are cosmetic reformatting (line-wrapping dict assignments and a blank line addition).

Notable pre-existing gaps (flagged in earlier review threads, not introduced by this PR):

• filename is still dropped on the reverse path (_transform_input_file_item_to_file_item in litellm/responses/litellm_completion_transformation/transformation.py) because its keys list only includes ["file_id", "file_data"].
• result.append(converted) is unconditional, so a malformed {"type": "file", "file": None} input silently produces an empty {"type": "input_file"} item.

Confidence Score: 4/5

• Safe to merge; the fix is correct, well-scoped, and tested with mock-only unit tests.
• The core bug fix is straightforward and correct — adding the missing elif item_type == "file" branch properly flattens nested file content. The two pre-existing edge cases (silent empty input_file when file=None, and filename loss on the round-trip) were flagged in earlier review threads and are not regressions introduced by this PR. No new logic paths are touched beyond the narrow fix.
• No files require special attention; the only file with a meaningful logic change is litellm/completion_extras/litellm_responses_transformation/transformation.py, and the change is small and well-tested.

Important Files Changed

Filename	Overview
litellm/completion_extras/litellm_responses_transformation/transformation.py	Adds explicit elif item_type == "file" branch in _convert_content_to_responses_format to map Chat Completion file items to Responses API input_file format, correctly flattening the nested file dict. Also contains cosmetic formatting-only changes to other parts of the file.
tests/test_litellm/completion_extras/litellm_responses_transformation/test_completion_extras_litellm_responses_transformation_transformation.py	Adds two mock-only unit tests verifying the new file → input_file conversion: one for the file_data+filename variant and one for the file_id variant. No real network calls; tests are correctly placed in the mock-only test folder.

Sequence Diagram

sequenceDiagram
    participant Caller as Caller (/chat/completions)
    participant FwdTransform as LiteLLMResponsesTransformationHandler<br/>_convert_content_to_responses_format
    participant ResponsesAPI as Responses API
    participant RevTransform as LiteLLMCompletionResponsesConfig<br/>_transform_input_file_item_to_file_item

    Caller->>FwdTransform: {type: "file", file: {file_data: "...", filename: "test.pdf"}}
    Note over FwdTransform: NEW: elif item_type == "file"<br/>Extract file_id/file_data/filename from nested file dict
    FwdTransform->>ResponsesAPI: {type: "input_file", file_data: "...", filename: "test.pdf"}

    ResponsesAPI-->>RevTransform: {type: "input_file", file_data: "...", filename: "test.pdf"}
    Note over RevTransform: keys = ["file_id", "file_data"]<br/>⚠ "filename" is NOT in keys list (pre-existing gap)
    RevTransform-->>Caller: {type: "file", file: {file_data: "..."}}
    Note over Caller: ⚠ filename is silently dropped on round-trip

Loading

_{Last reviewed commit: 98890e7}