Add Causal Chain Visualization for LLM Trace Debugging

[Aki-07]

Describe the feature or potential improvement

Add a "Causal Chain" visualization in LangFuse to link spans by input-output dependency. While LangFuse provides excellent trace inspection, developers currently have no way to see which upstream call or retrieved chunk led to a failure or hallucination. Understanding this causal link is crucial for debugging multi-step agent workflows (LangGraph, LangChain, DSPy, etc.).

Proposed Solution

• Extend span schema with optional parent_output_ref.
• Build a DAG-based view showing "causes" and "effects" across spans.
• Highlight failure propagation paths (e.g., red edges).
• Include summary metrics like failure_depth or dependency_density.

Additional information

I’d be glad to work on this feature end-to-end (schema changes + UI visualization). If this fits within the current roadmap, I can begin with a small prototype to validate the approach.

[Aki-07]

@hassiebp thoughts or comments?

[hassiebp]

fyi @nimarb

[nimarb]

Hi @Aki-07 , sounds very interesting ! Do you already have some sort of visualization in mind / vibed a prototype? Expanding the span schema is very delicate, but i'd be interested in how the output would look like (even drawn on a piece of paper) before we tackle this!

[Aki-07]

Thanks for reply @nimarb and your interest!, have rough sketched something. Would love your thoughts on this

[Aki-07]

@nimarb any thoughts???