Hi HN, I built llmSHAP, a multi-threaded open-source Python library that attributes how much different parts of LLM input contributed to the LLM’s output, using Shapley values.

It’s designed for LLM prompts / RAG contexts where you want something more structured than token masking: you can define “features” as tokens, sentences, paragraphs, tools, images, or arbitrary fields, pin “permanent” context that should always be included (e.g., system prompt + question), and compute attributions with options like caching and multi-threading to speed up the many model calls.

Repo + tutorial are here: https://github.com/filipnaudot/llmSHAP https://filipnaudot.github.io/llmSHAP/tutorial.html

I’d love feedback on if you think this could be genuinely useful in your LLM workflows (prompt engineering, RAG debugging, evals, guardrail auditing, etc.)?

Hi Filip,

Cool project! I’ve used TokenSHAP but I actually found llmSHAP via the preprint on arXiv a few weeks ago and found it to be a good alternative.

Are you planning on adding in-image object attribution?