LlamaIndex

An early RAG system built with naive vector search was producing hallucinated answers 28% of the time — unacceptable for compliance-critical product documentation queries.

Redesigned the retrieval layer using LlamaIndex's hybrid retriever combining dense vectors and BM25 keyword search. Added a re-ranking step with sentence-window retrieval to improve context quality before generation.

Hallucination rate dropped from 28% to 7.6% — a 73% reduction. Answer grounding scores (measured with RAGAS) improved from 0.61 to 0.94. System now serves 50k employees globally.

5,000 policy analysts were manually searching through 40 years of regulatory guidance documents, a process taking 3–5 hours per research task with inconsistent results.

Built a LlamaIndex document hierarchy with recursive summarization indexing — chapters, sections, and paragraphs all separately indexed with cross-references. Used metadata filtering to scope searches by regulation type, year, and jurisdiction.

Research time reduced 68%. A domain-expert evaluation panel validated answer accuracy at 96.3%. The system processes 12,000 queries per month with zero PII leakage.

A legal tech company needed to ingest 15 different document types (PDFs, Word, HTML, email threads, court transcripts) with automatic chunking, metadata extraction, and real-time updates as source documents changed.

Built a LlamaIndex ingestion pipeline using custom node parsers for each document type. Implemented incremental indexing with change detection so only modified documents re-index, cutting update latency from 4 hours to under 5 minutes.

Supports 50k concurrent users at <80ms search latency. 99.98% uptime over 12 months. Processing cost reduced 60% vs the previous batch-ingestion approach.