RAG architecture for financial data
Standard retrieval-augmented generation retrieves text and lets the model compose an answer. That design is unsafe for reported financial figures, because retrieval quality and model composition both sit between the ledger and the answer.
A finance-grade architecture separates three layers: a reconciled data layer, a deterministic calculation layer, and a language layer that explains results.
Retrieval serves context and narrative. It never serves arithmetic.
Questions and answers
What is different about RAG for financial data?
Financial answers must be reproducible and traceable. Retrieval alone cannot guarantee either, so arithmetic is moved out of the model into deterministic code that runs against a reconciled data layer, and retrieval is used for context and narrative.
What belongs in the retrieval layer for finance?
Policy documents, prior variance commentary, close checklists, contract and memo text, and internal definitions. These are language artefacts where retrieval adds real value.
What must not come from retrieval?
Any reported figure. Balances, consolidations, currency conversions, and variances are computed deterministically from reconciled source data, not assembled from retrieved snippets.
How is reconciliation handled before retrieval?
Source data is mapped to the customer's chart of accounts and entity structure and reconciled first. Retrieval and generation run on top of a reconciled layer, so a figure cannot come from an unmapped or stale extract.
How does lineage work in this architecture?
Every figure carries line-item lineage back to the general ledger, and every action is logged with who, what, and when. A reviewer can trace an answer to its source records without reading model internals.
Is a knowledge graph needed as well as a vector index?
They solve different problems. A vector index finds similar language; a graph holds the relationships between entities, accounts, periods, owners, and prior explanations. Finance workflows depend heavily on those relationships, so the graph carries the institutional memory and the index supports language retrieval.
Proof
- Reference client: $523K total contract value over 3 years.
- Approximately 430 analyst hours saved.
- Forecast cycle cut from about 60 hours to under 6 hours (90% faster).
- Error rate below 0.1%.
- 2,500+ autonomous tasks completed in month one.