Built a local legal clerk on top of CourtListener

[mpersinger87]

I’ve been building a local case-retrieval tool using the CourtListener corpus.

The goal wasn’t “AI lawyer.” It was something much simpler: a fast legal clerk that can help pull relevant opinions and useful excerpts without bouncing everything through cloud tools.

A few things I learned:
• it works a lot better when the query uses quoted doctrinal terms of art
• examples: "direct benefits estoppel" or "eight corners" "extrinsic evidence" Texas
• broad natural-language searches can drift
• exact legal phrases matter a lot
• retrieval quality lives and dies on metadata, chunking, and search discipline

What it does well right now:
• fast retrieval from a local corpus
• useful excerpts from matched opinions
• good results when the search is phrased like actual legal research

What still needs work:
• better metadata consistency
• stronger jurisdiction filtering
• better ranking on messy queries

I built it because I have a background in legal research and wanted something that behaves more like a clerk than a chatbot.

If it’s useful to people here, I’m happy to post more about what worked, what broke, and what I learned building it.

[robertkarl]

I'm interested in your work here. I'm working on on-premises AI with local LLMs/local hardware, and solving some of the pain points you mentioned.

[mpersinger87]

Very cool! Anything in particular that got you stuck? Maybe I can shed some light on it.