RAG

Language models work best when the information they need is already in the conversation. The challenge is that most organisations hold their most valuable knowledge in documents, databases, and codebases far too large to fit in any context window. RAG — Retrieval-Augmented Generation — is the architecture that bridges that gap: finding and surfacing only what is relevant, at the point it is needed.

Umage builds RAG systems as first-class tools for AI agents. The approach draws on deep experience with search — vector search layered with filters and facets to narrow retrieval as precisely as possible before anything reaches the context window. The underlying store varies by data type: relational databases like Postgres and SQL Server, NoSQL stores like Elasticsearch, or custom indexes built for the task. Chunking and embedding strategies are chosen per project, sometimes with open-source tooling, sometimes with approaches developed in-house. One recurring application is code intelligence: Umage has built RAG indexes that extract design patterns and API knowledge from large collections of NuGet packages, and uses tools like GitNexus and Chunkhound to support deep research in existing codebases — the same infrastructure that powers their own development workflow.