Skip to content

Title LiteLlm: Azure OpenAI file_id with assistant- prefix bypasses Responses-API content block #5664

Closed

Description

@yparwani

🔴 Required Information

Describe the Bug:

google.adk.models.lite_llm._looks_like_openai_file_id recognizes only the file- prefix as a valid OpenAI/Azure file id. Azure OpenAI's Files API actually returns ids prefixed with assistant- for purpose="assistants" uploads, and those ids are legitimately accepted by the Azure Responses API in the {"type":"file","file":{"file_id":"..."}} content block.

Because the prefix gate rejects assistant-..., ADK's content builder skips the Responses-API branch in _get_content (lines ~1095-1104 of models/lite_llm.py) and falls through to a later branch that emits a non-conforming content part of shape {"type":"file","file_uri":"assistant-...","mime_type":"application/pdf"}. Azure ignores this shape.

The same gate is referenced from _is_unsupported_file_uri and the redact helper (lines ~302 and ~322), so the impact is consistent: any caller that legitimately uploads to Azure OpenAI Files with purpose="assistants" (the standard purpose for PDF attachments to the Responses API) cannot send those files through ADK's LiteLlm

Steps to Reproduce:

  1. pip install google-adk litellm google-genai
  2. Upload a PDF to Azure OpenAI Files with purpose="assistants" via LiteLLM (litellm.acreate_file(file=..., purpose="assistants", custom_llm_provider="azure", api_key=..., api_base=...)). The returned id begins with assistant- (e.g. assistant-934058239058234095834).
  3. Build a genai_types.Part(file_data=FileData(file_uri="assistant-...", mime_type="application/pdf")) and feed it into an LlmRequest whose model is azure/<deployment>.
  4. Invoke LiteLlm and inspect the OpenAI-compatible message body LiteLLM sends to Azure (e.g. litellm.set_verbose = True or capture the messages list via a before_model_callback).

Expected Behavior:

The file content block matches the Azure Responses API contract:

{"type": "file", "file": {"file_id": "assistant-8b6ytv2iLxqCWtaMcxHSms"}}

i.e. assistant- ids are treated equivalently to file- ids when provider in {"openai", "azure"}.

Observed Behavior:

The content block is emitted as:

{"type": "file", "file_uri": "assistant-8b6ytv2iLxqCWtaMcxHSms", "mime_type": "application/pdf"}

file_uri and mime_type are not documented keys on a file-typed OpenAI/Azure content part. Azure rejects/ignores this attachment shape and the model never receives the file.

Environment Details:

  • ADK Library Version (pip show google-adk): 1.32.0
  • Desktop OS: macOS (also reproduces on Linux)
  • Python Version (python -V): 3.12

Model Information:

  • Are you using LiteLLM: Yes
  • Which model is being used: azure/gpt-5.4-mini (Azure OpenAI Responses API)

🟡 Optional Information

Regression:

N/A — _looks_like_openai_file_id has only ever accepted the file- prefix in releases I've inspected. This is a missing-case bug rather than a regression.

Logs:

N/A

Relevant ADK source (google/adk/models/lite_llm.py, ADK 1.32.0):

# line 282
def _looks_like_openai_file_id(file_uri: str) -> bool:
    """Returns True when file_uri resembles an OpenAI/Azure file id."""
    return file_uri.startswith("file-")

# line 1095
elif part.file_data and part.file_data.file_uri:
    if (
        provider in _FILE_ID_REQUIRED_PROVIDERS
        and _looks_like_openai_file_id(part.file_data.file_uri)
    ):
        content_objects.append({
            "type": "file",
            "file": {"file_id": part.file_data.file_uri},
        })
        continue
    # ... falls through to a non-Responses-API shape for assistant- ids

Screenshots / Video:

N/A.

Additional Context:

  • Azure OpenAI Files API behavior: uploads issued with purpose="assistants" (the only purpose value LiteLLM 1.83.x's acreate_file accepts that maps to a PDF attachment for the Responses API) return ids prefixed with assistant-. These ids are accepted by the Responses API {"type":"file","file":{"file_id":...}} content block.
  • Suggested fix: broaden the prefix check to file_uri.startswith(("file-", "assistant-")). All three internal call sites use the bare module-level name, so a one-line fix in _looks_like_openai_file_id propagates to _is_unsupported_file_uri, the redact helper, and the _get_content Responses-API branch.
  • Related observation: even if a caller wanted to inject the file id explicitly via extra_body, LlmRequest.config (a google.genai.types.GenerateContentConfig) does not expose an extra_body field, so the only correct paths today go through Part.file_data.file_uri → ADK's content builder. That makes the prefix gate the only thing standing between the caller and a correct payload.

Minimal Reproduction Code:

import asyncio
from google.adk.models import lite_llm as adk_lite_llm
from google.genai import types as genai_types

parts = [
    genai_types.Part(
        file_data=genai_types.FileData(
            file_uri="assistant-3402983204598",  # real shape from Azure OpenAI Files
            mime_type="application/pdf",
        )
    ),
]

content = asyncio.run(
    adk_lite_llm._get_content(parts, provider="azure", model="azure/gpt-5.4-mini")
)
print(content)
# Expected: [{"type": "file", "file": {"file_id": "assistant-..."}}]
# Actual:   [{"type": "file", "file_uri": "assistant-...", "mime_type": "application/pdf"}]

How often has this issue occurred?:

  • Always (100%) — deterministic for any Azure OpenAI Files upload with purpose="assistants".

Activity

  1. added
    models[Component] This issue is related to model support
    on May 12, 2026
  2. rohityan commented on May 12, 2026

    @rohityan
    Collaborator

    Hi @xuanyang15 , PTAL

  3. xuanyang15 commented on May 12, 2026

    @xuanyang15
    Collaborator

    @GWeale Could you please take a look?

  4. added a commit that references this issue on Jun 2, 2026
    b73679e
  5. added 2 commits that reference this issue on Jun 17, 2026
    796964a
    ea325c7
  6. added 3 commits that reference this issue on Aug 11, 2026
    15e8947
    9e57388
    0714d9b
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

models[Component] This issue is related to model support

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions