Skip to content

Insights / On-prem LLM

RAG on private data: the parts that actually matter

Retrieval-augmented generation on private corpora succeeds or fails on the parts of the system nobody puts in the architecture diagram.

Retrieval is the bottleneck

Most RAG systems on private data fail at retrieval, not at generation. The bottleneck is indexing the right things, retrieving the right things, and evaluating both. Generation quality is usually downstream of retrieval quality.

Evaluation is the unsexy part that wins

A repeatable evaluation suite on representative queries beats every prompt-engineering iteration. Without an eval suite, every change is a vibes-based release.

Access controls are part of the index

Document-level access controls have to survive the indexing pipeline. If the user-facing app respects ACLs but the index does not, the system leaks. The fix lands at the index layer, not at the response layer.

Start with the assessment.

A fixed-scope AI readiness assessment: your workflows, your data, your highest-ROI agent use cases, and a deployment roadmap. Two to four weeks.