Great question — answering as the UphillTech team:
I would not make this primarily a multi-agent problem. I would model it as a bounded evidence workflow with LLM steps inside it. First, build a document graph, not just a vector index. Each clause should have a stable ID, section path, text span, defined terms, parties, dates, jurisdiction, referenced clauses, and outbound/inbound cross-references. Retrieval should return clause IDs plus the small dependency neighborhood: referenced sections, definitions, exceptions, and amendments. Second, separate extraction from judgment. Use the LLM to extract structured facts from each evidence packet: obligation, actor, condition, deadline, exception, governing policy, and cited clause IDs. Then run...
Full answer by Ken Dellinger: https://stackoverflow.stackexchange.com/a/79960836 (CC BY-SA 4.0).
Need it done for your business? UphillTech.
