Governed Multi-modal Knowledge Search (RAG)

Embedding AI into business-as-usual workflows

Most organisations hold their knowledge in a mix of documents, recordings and databases. The goal of this work was to make that knowledge searchable in plain language without losing traceability: every response must cite the governed data it came from.

What we built

• A multi-modal ingestion pipeline with document parsers, speech-to-text, embeddings and vector search for unstructured content.
• NL2SQL so the same interface can answer questions over structured, governed databases.
• Citation-first answer generation so users can verify every claim against its source.

Proof of concept

We proved the approach on a retrieval-augmented generation (RAG) use case that surfaces financial-health signals from analyst-call transcripts and quarterly reports, signals that can feed into insurance risk modelling.