Need per-API / per-trace control to disable media uploads

[amaann2]

We’re using Langfuse via LangChain callback handlers (langfuse==2.60.8, langfuse.callback.CallbackHandler).

We came across an existing approach that allows disabling media uploads globally (for example, discussed here: https://github.com/orgs/langfuse/discussions/6179). That solution is helpful, but it doesn’t fully cover our use case.

In our setup, media gets uploaded automatically through the LangChain callback flow, and what we’re looking for is a more conditional / selective control:
• For some API flows (e.g. large PDFs), we’d prefer not to upload media
• For other flows (e.g. images for debugging), media uploads are useful
• A single global on/off switch makes it hard to support both cases

We’re hoping there’s a supported way to control media uploads per API flow / per trace / or based on payload size, especially when using LangChain callbacks.

[jannikmaierhoefer]

Hi @amaann2, how do you integrate with Langfuse? Are you manually creating all observations using the Python SDK, or are you also using one of our native integrations? cc @inkeep

[amaann2]

Hi @jannikmaierhoefer , thanks for checking.

We’re integrating Langfuse via LangChain callback handlers

Details:
• Langfuse version: langfuse==2.60.8
• Integration: langfuse.callback.CallbackHandler
• Langfuse is automatically logging LLM inputs/outputs through the LangChain callback flow
• Media (PDFs/images/base64) gets uploaded implicitly when present in the LLM inputs/outputs

This makes it difficult to control media uploads per API flow or per payload size, since uploads happen automatically via the callback handler rather than explicit SDK calls.

Our requirement is programmatic control (per flow / per trace / size-based), not just a global on/off switch.
Happy to share more details if needed.

[jannikmaierhoefer]

Thanks for adding this here @amaann2!

[inkeep-ai-assistant[bot]]

Based on the available documentation, I don't see a built-in SDK flag or environment variable specifically designed to disable media uploads (PDFs/images/base64) in the Langfuse Python SDK.

The Python SDK supports several environment variables for controlling behavior, including LANGFUSE_OBSERVE_DECORATOR_IO_CAPTURE_ENABLED which can disable input/output capture for the @observe() decorator(1), but this doesn't specifically target media uploads.

The documentation mentions that capturing large inputs/outputs may add overhead and can be disabled per decorator using capture_input=False or capture_output=False parameters(1), but again this is a general input/output control rather than media-specific.

To help provide a more specific answer, could you clarify:

1. How are you currently instrumenting your application? (e.g., using @observe() decorators, context managers like start_as_current_observation(), or manual observations with start_observation())(1)
2. Are you using any native integrations like the OpenAI SDK integration(2), or is this purely custom instrumentation?
3. Where in your workflow are these PDFs being captured - as inputs, outputs, or metadata?

Understanding your integration method will help determine if there's a workaround using existing SDK features or if this truly requires a new feature addition.

📚 Sources:

1. Instrumentation
2. Observability for OpenAI SDK (Python)

---

Have another question? Just tag @inkeep.