Citation-Grounded RAG Is an Indexing Problem

A surprising amount of what fails in enterprise RAG systems is not the language model. It is the index.

We learned this the hard way while building Ravin, our contract intelligence platform. The first version relied on the obvious approach: chunk the documents, embed the chunks, retrieve the top-k. The model fluently answered questions. It also hallucinated clauses that didn't exist, missed clauses that did, and refused to commit to a citation.

What we changed

Three things. None of them were prompt engineering.

1. We treat documents as structured objects, not text streams. A contract has a clause hierarchy, defined terms, signatures, and schedules. We preserve that structure in the document model. The retriever becomes clause-aware.
2. We dual-index everything. Once for semantic similarity, once for exact lexical match. The model receives both signals and decides which to trust.
3. We cite by reference, not by location. A citation must point at a stable identifier — clause number, defined term, exhibit — that survives a re-render of the source document.

What this buys

The model now answers questions it could not before. More importantly, it answers them in a form a counsel can stand behind. Every response carries a citation the operator can open in the source PDF.

This is unglamorous work. There is no headline metric. There is, however, a single lawyer in a meeting who stops interrupting the demo to check the citation — and that is the metric we care about.