RAG and Retrieval-Augmented Generation

Key takeaways

• RAG is a data product: retrieval quality comes from source selection, chunking, indexing, ranking, and evaluation, not just the model.
• The pipeline runs source inventory, chunk and normalize, index, query rewrite, permission filter, retrieve and rerank, then generate with citations.
• Filter retrieval by user and tenant permissions so the system never surfaces documents the user could not open directly.
• Evaluate recall, precision, groundedness, and usefulness, and add a low-confidence fallback rather than checking model output alone.

RAG is a data product. Retrieval quality depends on source selection, chunking, indexing, permission filtering, ranking, and evaluation, not only on the model.

Pipeline

Quality Controls

• Track source freshness and ownership.
• Filter retrieval by user and tenant permissions.
• Prefer citations or source links for factual answers.
• Evaluate recall, precision, groundedness, and usefulness.
• Add a fallback when confidence is low.

Red Flags

• The system retrieves documents the user could not open directly.
• Old content outranks current policy.
• Evaluation only checks model output, not retrieval quality.