How does Alex OS use RAG?

Alex OS answers questions about this site by retrieving from the site's own content rather than from model memory. Published CMS content is chunked and embedded into a pgvector store; a question is embedded the same way, matched against those vectors, filtered by a per-mode relevance floor, and the surviving passages are passed to the language model as the only source it is allowed to answer from.

The retrieval layer is tuned for a small, high-signal corpus rather than a large public one: it detects when a question names a content type and widens the result set for browsing, boosts matches on titles and type affinity, and demotes structural navigation nodes so a page's menu label cannot outrank its substance. When nothing clears the relevance floor, the assistant says so instead of answering from the model's own priors.

Last verified 28 July 2026