DEV Community

CourtGPT
CourtGPT

Posted on

Open Source Tools for Legal Research Automation in 2026

#opensource #legaltech #ai #rag

Open Source Tools for Legal Research Automation in 2026

Legal research automation uses a combination of retrieval, language models, and tooling. This article catalogs open-source tools that practitioners and developers can use without per-seat SaaS fees. Each tool has a documented license and adoption footprint.

Document corpora

  • CourtListener (https://courtlistener.com): Federal + state court decisions, RECAP project, public API. Free Research tier.
  • RECAP: Browser extension that pulls PACER documents. Community-contributed archive.
  • Library of Congress: Federal statutes, regulatory history, executive orders. Public domain.
  • State legislature websites: All 50 states publish session laws and statutes; many have structured data.
  • LawInsider: Contract corpus; some sections accessible for free, full access paid.
  • Harvard CAPAPI: Free case law + statutes API. https://case.law
  • Court Forms: Each court has a public forms library (PDF). Form autofill needs OCR + classification.

Retrieval and vector databases

  • Postgres + pgvector: Open-source vector extension. HNSW index, 1536+ dims. Operational simplicity.
  • OpenSearch: Apache 2.0. BM25 + vector (k-NN) + RRF built in. Production-grade.
  • Weaviate: Open-source + managed. Modular vector DB with multi-vector support.
  • Qdrant: Rust-based. Open-source + managed. Strong latency claims.
  • Meilisearch: Open-source. Hybrid BM25 + vector. Limited scaling for >10M docs.
  • Milvus: Open-source. Strong at billion-scale.

LLMs and inference

  • Ollama: Local inference server. Open-source.
  • llama.cpp: CPU-friendly. Wide model support.
  • vLLM: Open-source serving framework. Strong throughput.
  • Open WebUI: Self-hosted chat UI.
  • LiteLLM: API gateway that unifies OpenAI, Anthropic, etc.

Models worth evaluating:

  • Llama 3.3 70B: Open-source, function calling, 128K context
  • Mistral Large: Open-weight, strong reasoning
  • Qwen 2.5: Stronger for non-English languages
  • DeepSeek V3: Open-weight, 671B MoE

Embeddings

  • Sentence-Transformers (Python): On-host embeddings. Models: all-MiniLM-L6, BGE-large, Nomic embed.
  • Ollama embeddings: Embed via local models.
  • voyage-law-2: Domain-specific for legal. (Closed-source but free tier)
  • OpenAI text-embedding-3-large: 3072 dims. Paid.

Eval frameworks

  • ragas: RAG-specific eval. Open-source. https://docs.ragas.io/
  • TruLens: Eval + observability. Open-source. https://www.trulens.org/
  • DeepEval: Eval framework. Open-source.
  • Promptfoo: Eval + red-team. Open-source.
  • OpenAI evals: Open-source eval suite. Hundreds of pre-built eval templates.

Document processing

  • Apache Tika: Apache 2.0. PDF, DOCX, HTML extraction.
  • Unstructured.io: Apache 2.0. PDF/HTML cleaning.
  • docling: Apache 2.0. Multi-format document parsing.
  • Tesseract: Apache 2.0. OCR.
  • Google Cloud Document AI: Commercial OCR.

Workflow automation

  • n8n: Apache 2.0 (recent). Open-source. 400+ integrations.
  • Apache Airflow: Apache 2.0. Workflow orchestration.
  • Temporal: MIT-licensed. Durable workflows.

Use case patterns

Pattern A: Public self-help portal 

Web portal (open source: Hugo / Next.js)
  -> Query
  -> Self-hosted RAG (Ollama + pgvector + LangChain)
  -> Streaming response with citations
Enter fullscreen mode Exit fullscreen mode

Laws.AI, Texas Law Help, MassLegalHelp.org use similar architectures.

Pattern B: Internal research tool 

SSO (OpenID Connect) -> Auth gateway
  -> Lawyer-facing UI (Next.js)
  -> Self-hosted OpenSearch + pgvector
  -> BAA-protected LLM endpoint (Azure OpenAI / Bedrock)
  -> Audit log + citation verification
Enter fullscreen mode Exit fullscreen mode

Pattern C: Court clerk assistant 

On-prem deployment (no egress)
  -> Fine-tuned Llama 3.3 70B (QLoRA)
  -> Private corpus (state rules + local procedures)
  -> Strict access controls
  -> Disclosure to bench
Enter fullscreen mode Exit fullscreen mode

Compliance considerations

  • HIPAA: BAA-protected LLM endpoints only for health data workflows. AWS Bedrock, Azure OpenAI offer BAAs.
  • Privilege: Privileged data must NOT be sent to non-BAA endpoints.
  • Discovery: AI-generated evidence may require disclosure under FRCP 26.
  • Sanctions: Filing briefs with fabricated AI citations has led to $5K-$31K sanctions in multiple cases.

Cost: open-source vs commercial

For 100K queries/mo:

Stack Open-source Commercial (BAA)
Vector DB $0 (self-host) $70-700 (Pinecone/Weaviate)
LLM $1,500/mo (Llama 3.3 70B on 4x A100) $3,000-5,000/mo
Embedding $0 (local) $200-500/mo
Workflow $0 (n8n) $24-99/mo
Hosting $500-2,000/mo varies
Compliance manual $500-2000/mo (BAA fees)
Total $2,000-3,500/mo $4,000-8,000/mo

Open-source saves ~50% but adds compliance + ops overhead.

Adoption signal

As of Q1 2026:

  • CourtListener: 1M+ monthly users, public API.
  • pgvector: Production at 100+ companies.
  • Mistral / Llama: Real legal-research deployments in EU + US.
  • OpenSearch: Used by 5+ state courts for internal search.
  • n8n: 100K+ active deployments.

References

Acknowledgments

This article catalogs widely-adopted open-source projects as of early 2026. Versions evolve; check official documentation for the latest.

Dillon Deutsch has built legal research systems using these open-source tools. https://courtgpt.ai

Top comments (0)