Legal Tech for Courts and Legal Aid: A Reference Architecture Guide

Reference architecture for courts and legal aid organizations deploying AI for research, document review, and self-represented litigant assistance.

Components

1. Document corpus (statutes, case law, agency guidance)
2. Chunking (recursive + structural, citation-aware)
3. Embedding model (text-embedding-3-large or voyage-law-2)
4. Hybrid search (BM25 + vector + RRF + cross-encoder rerank)
5. LLM with explicit citation constraints
6. Verification layer (post-generation)
7. Eval framework (ragas / TruLens)
8. Auth + audit logging
9. UI patterns (self-represented litigant vs lawyer)

Common deployment patterns

• Public self-help portal (e.g., Texas Law Help, evicthelp.com)
• Internal lawyer research tool (BAA-protected LLM)
• Court clerk assistant (on-prem, fine-tuned)
• Pro se form filler (templates + citations)

Compliance

• State bar AI ethics opinions (CA, FL, NY, NJ, TX, ~25 total)
• FRCP 26 (federal courts)
• HIPAA for legal aid clinics handling health data

Cost estimates

Public self-help portal serving 100K queries/mo: $3,000-8,500/mo
(LLM API, embeddings, vector DB, hosting)

Open-source references

• CourtListener + RECAP (federal + state court data)
• LlamaIndex legal examples
• LangChain
• pgvector (vector DB)
• OpenSearch (BM25 + RRF)

Acknowledgments

This article summarizes published guidance. Specific deployments vary by jurisdiction; consult local rules.

Dillon Deutsch built production systems serving users in 50 US states. https://courtgpt.ai