I have an LLM agent that processes documents and extracts statements and conditions. I want to evaluate how much of the document is being covered by the agent — specifically, whether any important statements or conditions are being missed.
Is there a good method or metric to assess this kind of coverage? Ideally, I’d like to quantify the completeness of the extraction with respect to the original document. Any suggestions on techniques, tools, or evaluation strategies would be appreciated.
If important, I am using Qwen3-4B and document in question is the RISC-V ISA Manual.