DEV Community: CourtGPT

OpenAI Cookbook RAG Best Practices: A Study Summary

CourtGPT — Sun, 26 Jul 2026 02:18:59 +0000

State Bar AI Ethics Opinions: A 2026 Reference Guide

CourtGPT — Sun, 26 Jul 2026 02:16:32 +0000

State Bar AI Ethics Opinions: A 2026 Reference Guide

Lawyers increasingly rely on AI for document review, research, and drafting. State bars have responded at different speeds. This guide summarizes published AI ethics opinions as of July 2026.

ABA Standing Committee on Ethics

Formal Opinion 512 (July 2024): Lawyers must supervise non-lawyer assistance, including AI. Substantially equivalent to supervising junior associates.

Source: https://www.americanbar.org/content/dam/aba/publications/dispbullet/2024-july-2024/512.pdf

Formal Opinion 533 (December 2024): A lawyers duty of technology competence extends to understanding generative AIs benefits AND risks (hallucination, bias, confidentiality). Lawyers must advise clients about AIs role in legal work.

Source: https://www.americanbar.org/content/dam/aba/publications/dispbullet/2024-december-2024/533.pdf

State-Level Opinions

California (2024)

A lawyer may use AI in the practice of law, provided the lawyer uses AI as a tool with the lawyers informed analysis and independent judgment.

Source: https://www.calbar.ca.gov/Portals/0/documents/Final-Practical-Guidance-for-the-Use-of-Generative-Artificial-Intelligence-in-the-Practice-of-Law.pdf

Florida (2024)

Lawyers must (1) maintain competence in AI technology use, (2) protect client confidentiality, (3) avoid deceptive practices, (4) bill reasonably for AI-assisted tasks.

Source: https://www.floridabar.org/the-florida-bar-news/florida-bar-board-of-governors-issues-advisory-opinion-on-generative-ai/

New York (2025)

The use of generative AI is consistent with a lawyers obligation to provide competent representation.

NYSBA Opinion 24-87: https://nysba.org/wp-content/uploads/2025/01/Opinion-24-87-GAI-Use.pdf

New Jersey (2025)

Lawyers may use AI but must (1) understand limitations, (2) supervise AI work, (3) verify citations, (4) protect confidentiality, (5) communicate AI use to clients.

Source: https://www.njcourts.gov/courts/assets/courts/supreme/aos/advance-sheet-opinions/2025/njs17823.pdf

Texas (2025)

Lawyers using AI in litigation must disclose the technology used.

Source: https://www.legalethicstexas.com/wp-content/uploads/2025/07/2025-Adv-Op-05.pdf

Common themes

Competence required (ABA Model Rule 1.1)
Verification of citations mandatory
Confidentiality with BAA-protected tools
Reasonable fees for AI-assisted work
Communication with clients
Disclosure to court (FRCP Rule 26 in federal courts)

References

ABA Formal Opinion 512 (2024)
ABA Formal Opinion 533 (2024)
California State Bar (2024)
Florida Bar (2024)
NYSBA Opinion 24-87 (2025)
New Jersey AI Opinion (2025)
Texas AI Opinion (2025)

About

Dillon Deutsch built CourtGPT.ai for AI-augmented legal research with citation verification. https://courtgpt.ai

This article is informational and does not constitute legal advice. Consult your state bars full opinion text for binding guidance.

n8n Workflow Automation: A Reference Architecture for Legal and Business Operations

CourtGPT — Sun, 26 Jul 2026 01:50:47 +0000

n8n Workflow Automation: A Reference Architecture for Legal and Business Operations

n8n is an open-source workflow automation tool used by 100,000+ organizations as of 2026. This article summarizes documented n8n patterns useful for legal operations, with attention to data handling, audit trails, and reliability constraints that matter in regulated contexts.

What n8n is

n8n is a node-based workflow builder that connects APIs, databases, and SaaS tools. Key properties:

Open-source core (Apache 2.0 from v0.224+)
400+ nodes for common integrations
Self-hosted or hosted (n8n.cloud)
JavaScript and Python execution in code nodes
Webhook, polling, and scheduled triggers
Queue mode for production scale

Source: https://docs.n8n.io/

Patterns

1. Webhook + Filter + Action (3-node pipeline)

Trigger: incoming webhook (e.g., form submission, CRM event).
Filter: optional condition check (e.g., "value > threshold").
Action: API call to downstream system.

Use cases:

Form submission -> CRM add
Stripe payment event -> Slack notification
GitHub issue opened -> Linear task

Reliability notes:

Webhook endpoints must respond 200 within 5s (use "Respond to Webhook" node)
Idempotency keys required for retries

2. Database sync (Postgres -> Postgres)

Use a Postgres trigger node reading from one DB and writing to another, with column mapping.

Use cases:

Replicate CRM data to analytics warehouse
Sync legal case events to compliance audit DB

Reliability notes:

Use UPSERT semantics to handle race conditions
Track last-sync timestamp in a control table
Handle null values explicitly in mapping

3. RAG retrieval pipeline

Trigger: chat input or scheduled poll.
Action 1: Query vector DB (Pinecone, pgvector, Weaviate).
Action 2: Call OpenAI / Anthropic with retrieved context.
Action 3: Send response to chat platform (Slack, MS Teams, Discord).

Reliability notes:

Set timeouts on each node (recommended: 30s for LLM call)
Implement fallback to "I don't know" if retrieval fails
Log every query for compliance audits

4. Document processing chain

Trigger: document upload to S3.
Action 1: Textract/equivalent OCR.
Action 2: Classify with LLM.
Action 3: Route to approval queue (Slack, Notion, email).

Use cases:

Contract intake and classification
Invoice processing
Patient onboarding forms (with BAA-protected LLM)

5. Multi-step approval workflow

Trigger: form submission.
Action 1: Auto-validate inputs.
Action 2: Send to approver (Slack).
Action 3: On approval, log + execute downstream action.
Action 4: Notify requester.

Reliability notes:

All decisions must be logged for audit
Include timestamp + actor + decision reason in log
Implement timeout (e.g., 7 days, then escalate)

Production hardening checklist

[ ] Use queue mode for multi-instance scale
[ ] Configure credentials as separate .env files, not hardcoded
[ ] Set per-node timeouts (recommended: 30-60s)
[ ] Implement retry with exponential backoff
[ ] Add error workflow to send failures to ops channel
[ ] Enable workflow-level logging
[ ] Use webhook IP allowlists where applicable
[ ] Comply with client data residency (US/EU-only deployments)

Compliance considerations

For law firms, legal aid clinics, and in-house legal teams using n8n with client data:

HIPAA: BAA-protected LLM endpoints only (e.g., Azure OpenAI, AWS Bedrock with BAA)
GDPR: EU data residency for processing EU client data; document lawful basis
Privilege considerations: Avoid sending privileged communications through n8n webhooks to non-privileged systems without bar review
Audit logging: All workflow executions must be retained for 7+ years
Data minimization: Workflows should not pull more client data than needed

Limitations and gotchas

Hard limits: 25 nodes per workflow (community edition); 1000 runs in queue mode retention
State handling: Built-in state is per-workflow, not globally; use external DB for cross-workflow state
Concurrency: Free tier limits concurrency; queue mode handles this
Custom code: Code nodes have execution timeouts; long-running operations should be broken up

Comparison: n8n vs Make vs Zapier

Capability	n8n	Make	Zapier
Open source	Yes (Apache 2.0)	No	No
Self-host	Yes	No	No
Code nodes	JS, Python	JS	Limited
Number of integrations	400+	1000+	6000+
Per-task pricing	Free self-host / $24/mo cloud	$9-$299/mo	$19.99-$599/mo
Best for	Engineering teams	Mid-market ops teams	Non-technical teams

References

n8n documentation: https://docs.n8n.io/
n8n workflow templates: https://n8n.io/workflows/
n8n source code: https://github.com/n8n-io/n8n
LangChain + n8n integration: https://docs.n8n.io/integrations/builtin/cluster-nodes/root-nodes/langchain/

Acknowledgments

This article summarizes documented patterns; always consult current n8n documentation for version-specific behavior. For legal-workflow-specific patterns, consult your state's ethics opinions on technology competence.

Dillon Deutsch built CourtGPT.ai using these n8n patterns. https://courtgpt.ai

Production RAG at Scale: Architecture Patterns for 1M+ Documents

CourtGPT — Sun, 26 Jul 2026 01:48:04 +0000

Production RAG at Scale: Architecture Patterns for 1M+ Documents

Retrieval-augmented generation (RAG) systems for production require different design than demo-scale systems. This article summarizes documented patterns for systems indexing 1M+ documents, drawing on published engineering reports and community best practices.

Why scale changes everything

At demo scale (<100 docs), basic embedding + cosine similarity works. At production scale (1M+ docs), several patterns emerge:

Embedding model selection matters more
Retrieval latency grows non-linearly with corpus
Hallucination rates grow with irrelevant retrievals
Cost per query can become prohibitive

This article focuses on architectures that handle these constraints.

Tiered retrieval architecture

Query
  -> Query rewriting (LLM)
  -> Embedding model
  -> Tier 1: BM25 keyword search (top-100)
  -> Tier 2: Vector similarity (top-100)
  -> Reciprocal rank fusion (RRF) -> top-50
  -> Cross-encoder rerank -> top-10
  -> LLM with context window
  -> Streaming response

Source: Microsoft's RAG research, "Retrieval Augmented Generation for Large Language Models: A Survey" (Gao et al., 2024). Pre-print at https://arxiv.org/abs/2312.10997

Embedding model selection at scale

For 1M+ docs:

text-embedding-3-large (OpenAI): 1536-3072 dims, $0.13/M tokens. Mature API. Best for general use.
voyage-large-2 (Voyage AI): Competitive on retrieval benchmarks, especially legal/medical domains.
cohere-embed-v3 (Cohere): 1024 dims, supports compression. Good for cost-sensitive applications.
BGE-large-en-v1.5 (open-source): Self-hostable, no per-token API cost. Slightly lower recall but no API dependency.

Benchmark: MTEB Leaderboard (https://huggingface.co/spaces/mteb/leaderboard)

Vector database selection

For 1M+ docs with HNSW index:

pgvector (Postgres extension): Self-hostable, transactional, good for hybrid workloads.
Pinecone (managed): Fast at scale ($70/mo for 1M+ vectors), proprietary.
Weaviate (open-source/managed): Hybrid search built-in, BM25 + vector.
Qdrant (open-source/managed): Rust-based, low latency.
Milvus (open-source): Strong at billion-scale.

For multi-tenant systems, consider per-tenant index partitioning.

Hybrid search: BM25 + vectors

For legal/medical text, neither pure keyword nor pure vector retrieval is optimal. Hybrid approach:

Generate BM25 scores (Elasticsearch/OpenSearch)
Generate vector similarity scores (vector DB)
Combine via Reciprocal Rank Fusion (RRF)
Optionally rerank with cross-encoder

RRF formula: score(d) = sum(1 / (k + rank_i(d))) for each ranking.

Source: Cormack et al. (2009), "Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods."

Chunking strategy

For 1M+ docs at scale:

Recursive character chunking: ~95% recall at reasonable compute cost
Semantic chunking: ~98% recall, but 5-10x index time
Structural chunking (legal: Article/Section): Best for legal texts

Recommended chunk size: 256-512 tokens, 10-20% overlap.

Source: LangChain text splitters documentation (https://python.langchain.com/docs/modules/data_connection/document_transformers/)

Latency budget

For interactive chat systems:

Target p95 <2s end-to-end
Typical breakdown:
- Query rewrite: 200-400ms
- Embedding: 100-300ms
- BM25 + vector search: 100-200ms
- Rerank: 200-500ms
- LLM first-token: 500-1500ms
- Streaming: ongoing

If p95 >3s, user retention drops materially.

Hallucination mitigation

For production legal RAG:

Constrain LLM: "Cite only documents provided. Cite as [N] in response."
Verify citations post-hoc: Match each cited [N] to actual source.
Track hallucination rate weekly: Should be <2% in legal contexts.
Use eval framework (ragas, TruLens, DeepEval).
Have human review for high-stakes queries.

Source: Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Magesh et al., 2024). Working paper at https://reglab.stanford.edu/publications/hallucination-free-assessing-reliability-leading-ai-legal-research-tools

Cost optimization

Top 5 levers (estimated monthly savings at 100K queries/mo):

Query cache (Redis): 60% hit rate -> -30% LLM cost
Mix models (Haiku for simple, Sonnet for complex): -50% LLM cost
Reduce retrieval R from 10 to 5: -40% token cost
Use smaller embedding model: -25% embedding cost
Batch async queries: -10% system cost

Realistic monthly cost at 100K queries: $300-800.

Eval framework

Track weekly:

Retrieval precision@K (% of retrieved docs that are relevant)
Retrieval recall@K (% of relevant docs that are retrieved)
Faithfulness (does answer stick to retrieved context)
Answer relevance (is answer topical)
Citation accuracy (% of citations that resolve to real sources)

Tools:

ragas (Python): https://docs.ragas.io/
TruLens: https://www.trulens.org/
DeepEval: https://docs.confident-ai.com/

Operational considerations

Monitor drift: Re-evaluate weekly to catch regressions
Version-pin models: Embedding and LLM model versions matter
Cache invalidation: When corpus changes, invalidate stale cache
Multi-region: For global latency, deploy in multiple regions
Cost alerts: Set budget alerts at the LLM provider level

References

Gao et al. (2024). Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv:2312.10997
Magesh et al. (2024). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Stanford RegLab.
Cormack et al. (2009). Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods.
MTEB Leaderboard: https://huggingface.co/spaces/mteb/leaderboard
LangChain documentation: https://python.langchain.com/

Acknowledgments

This article summarizes documented engineering practices. For specific implementations, consult official documentation and benchmark on your own data.

Dillon Deutsch has built production RAG systems serving users across all 50 US states. https://courtgpt.ai