Reference documentation for ScaleCloud’s standard retrieval-augmented generation system architecture, covering the retrieval pipeline components, evaluation methodology, and the access-control pattern applied at the retrieval layer.
Pipeline components
The standard pipeline includes a document ingestion and chunking stage, an embedding stage, a hybrid (vector plus keyword) retrieval stage, and a generation stage with mandatory source citation.
Access control at retrieval time
Document-level permissions are always enforced during retrieval, before any content reaches the language model — never as a post-generation filter on the model’s output, which cannot reliably prevent a permission leak.
Leave a Reply