Passive retrieval performs a vector search for every user query, whether external evidence is needed or not. It connects an LLM’s parametric memory with non-parametric memory in an external vector database, making retrieval a fixed baseline rather than a query-dependent decision. section 9.0
Its indiscriminate behavior creates several production tradeoffs:
- Wastes compute on queries that do not require retrieval. section 9.0
- Pollutes context windows with irrelevant passages that crowd out useful answers. section 9.0
- Can leave generation ungrounded despite presenting retrieval as protection against fabricated citations. section 9.0
Production baseline
A passive RAG service receives a user query, runs vector search automatically, and places the returned passages into the generation context. Adaptive systems such as Self-RAG instead retrieve on demand, may retrieve multiple times during generation, or skip retrieval entirely. section 9.0
