LLM
EpsteinRAG – RAG Search Engine
Full-stack RAG search engine enabling natural language querying across 44,886+ declassified government documents with AI-generated answers and source citations.
FastAPINext.jsPostgreSQLMeilisearchGeminiPython
System Architecture
[SYS] ─── epsteinrag-hybrid.arch
Overview
A full-stack RAG search engine for natural language querying across 44,886+ declassified documents.
Problem Statement
Declassified documents are difficult to search using traditional keyword methods.
Solution & Approach
Hybrid search: keyword (Meilisearch) + semantic vector search (Gemini Embedding 001, 3072-dim) with pgvector.
Key Features
- Hybrid search: keyword + semantic vector (3072-dim embeddings)
- ~16ms cached responses via SHA-256 query caching
- Real-time streaming via Server-Sent Events (SSE)
- Reduced latency from 8-10s to 2-3s for first token
- Citations linked to original DOJ documents
Results & Impact
~16ms cached responses with streaming. All answers include numbered citations.
Lessons Learned
Hybrid search combining keyword and semantic provides better results than either alone.