AntflyAntfly
Back to Builder Stories

Epstein Search

AI-powered search across millions of publicly released court documents and DOJ files, delivering cited answers where other RAG pipelines return nothing.

Millions
Documents indexed
15K+
Pages searchable
cited
AI answers with sources
hybrid
Semantic + full-text search
Epstein Search

What it is#

Epstein Search is an AI-powered search engine that indexes over 4,000 publicly released court documents and DOJ files from the Jeffrey Epstein investigation. Users can ask natural language questions and get AI-generated answers with inline citations pointing to specific pages in the source documents.

Ask "What is Great St. Jim, LLC?" and you get a high-confidence answer explaining it's a Virgin Islands LLC that acquired parcels on Great St. James Island, complete with citations from page 9109 of the DOJ dataset:

High-confidence answer for Great St. Jim, LLC

Where other RAG pipelines fail#

The same publicly available documents power several search tools. But most of them struggle to return meaningful results for even straightforward questions.

Epstein Archive (epstein-docs.github.io) indexes 8,175 documents but returns "No results found" for "What is Great St. Jim, LLC?":

Epstein Archive returning no results

UnCovered (uncovered.today) can't find information about who visited Little St. James Island -- one of the most well-documented topics in the files:

UnCovered returning no information

Epstein Search answers both questions with cited sources, confidence scores, and follow-up questions to dig deeper:

Cited answer with follow-up questions

How it works#

The pipeline starts with PDF ingestion -- thousands of court filings, depositions, and DOJ files are parsed and split into chunks. Each chunk is stored in Antfly with both a full-text BM25 index for keyword matching and an AKNN vector index for semantic similarity.

When a user asks a question, both indexes are queried simultaneously and results are merged using reciprocal rank fusion. This hybrid approach means exact legal terms and entity names get matched by BM25, while semantically related passages are surfaced by the vector index -- even when the phrasing doesn't match.

SearchAF then synthesizes the top results into an AI-generated answer with inline citations, a confidence score, and suggested follow-up questions.

Stack#

  • Search & Storage: Antfly hybrid search (BM25 + AKNN vector)
  • AI Answers: SearchAF with AI-generated answers, citations, and follow-up questions
  • Embeddings: Antfly Inference for vector encoding of document chunks
  • Source Data: millions of publicly released court documents and DOJ files
Architecture
1
PDF Ingestion
Thousands of publicly released court documents and DOJ files are parsed and chunked into searchable segments
2
Hybrid Indexing
Each chunk is indexed with both BM25 full-text and AKNN vector indexes in Antfly
3
Semantic Search
Queries hit both indexes simultaneously, results fused with reciprocal rank fusion
4
AI Answers
SearchAF generates AI answers with inline citations, confidence scores, and follow-up questions