Onyx + Claude Code: Docker RAG setup and comparisons

The tutorial compares five approaches to asking questions about your own documents: Claude chat, Claude Projects, NotebookLM, Onyx and a custom RAG build. The speaker recommends moving up only when simpler options stop meeting your needs. His signals for RAG include repeated searches across a large corpus, verifiable citations, frequently changing data and a product that needs retrieval.

The self-hosted walkthrough uses Onyx with Docker on a MacBook Pro, with Claude Code helping install and manage it. The speaker reports that his setup needed a 14 GB Docker memory allocation. He distinguishes the full installation from the light version, which lacks the RAG components, and explains that Onyx needs a separately configured model provider. His demo uses Claude through an API; he says a local LLM is also an option. Hosting Onyx yourself therefore does not establish that all processing stays on your hardware.

Data preparation covers an inventory, deduplication, sensitive information removal and document sets. The testing section examines split chunks, lost context, missed exact matches and unsupported answers. Claude generates 12 questions covering retrieval, cross-document answers, faithfulness and abstention. An initial score of 7/12 becomes 12/12 after review, a result specific to the synthetic demo data. The speaker calls for human checks and weekly evaluation reruns as documents change, and notes that some user-management features are edition-dependent.