Retrieval-Augmented Generation

How RAGVue Legal transforms queries into accurate, cited responses

Step 1: Data Curation (The Foundation)

The Hidden Work: Most RAG failures happen before the first query. Garbage in, garbage out. Quality outcomes require structured, curated data - not raw document dumps. This is the skill that separates professional RAG from toy demos.

1
Raw Data
PDFs, HTML, scanned documents with inconsistent formatting, headers, footers, page numbers
Unusable for AI
2
Clean & Extract
Remove white noise: headers, footers, page numbers, formatting artifacts, duplicate content
Plain text
3
️
Structure
Parse into logical units: sections, citations, cross-references. Add metadata: jurisdiction, date, type
Structured data
4
Embed & Index
Generate vector embeddings, create search indexes, validate relationships
Ready for RAG
200+
Hours of Curation
99.84%
Clean Parse Rate
157K
Documents Processed
Step 2: The RAG Pipeline (Query Time)
1
User Query
Natural language question about legal matter
2
Embed Query
Convert text to 384-dimension vector
3
Vector Search
Find semantically similar documents
4
Retrieve Context
Pull relevant statutes & cases
5
LLM Generation
Generate response with citations
Curated Knowledge Base

PostgreSQL + pgvector

157,000+ embedded legal documents with hybrid search

WA RCW
US Code
US Constitution
Federalist Papers
WA Constitution
Court Opinions
Why RAG Matters

❌ Without RAG (Raw LLM)

  • Trained on stale data (knowledge cutoff)
  • Hallucinations - confident but wrong
  • No citations or sources
  • Generic, non-jurisdictional answers
  • Can't access your private data

✅ With RAG (RAGVue)

  • Current, curated legal corpus
  • Grounded in retrieved documents
  • Every claim has a citation
  • Jurisdiction-specific results
  • Works with your private data
Air-Gapped Architecture
️
Local Embeddings
sentence-transformers
runs in PostgreSQL
PostgreSQL 18
pgvector + plpython3u
All logic in database
Ollama (Local)
Llama 3.1 on localhost
No API calls
Audit Trail
Immutable logs
Hash chain verified