HLI-RAG
A retrieval engine that indexes lazily, re-ranks in parallel and checks answers against the source.
AI engine / 2025—26
HLI-RAG.
A retrieval engine that indexes lazily, re-ranks in parallel and checks answers against the source.
THE BUILD / AI engine
HLI-RAG is a retrieval engine built to stop confident nonsense. It indexes lazily, re-ranks candidates in parallel and runs a natural-language inference check before an answer is allowed out.
The pipeline
- Lazy ingest with content hashing so unchanged documents are skipped
- Hybrid retrieval: dense vectors plus keyword search
- Parallel re-ranking with MMR for diversity
- NLI check between answer and source chunks
- Citation assembly with exact spans
Why it exists
Most RAG demos answer beautifully and cite nothing. This one would rather say “I don’t know” than invent a source.
NEXT / YOUR PROJECT
WANT ONE
LIKE THIS?
Tell me what it should do. I’ll reply with a plan and a real price.
START A BRIEF ↗