Research, RAG and Codex workflows — how should sources become useful without polluting the manuscript?

[qnbs]

Why this exists

Research can enrich a project without becoming indistinguishable from the author's manuscript. This Discussion explores how Capture Inbox material, source libraries, retrieval, and Codex-style workflows can remain useful, inspectable, and reversible.

Product decision space

Consider a path from Capture Inbox material into a research library containing URLs, PDFs, notes, images, and excerpts as untrusted sources. Organization and tagging should preserve source provenance and citations. Project-local semantic search and transparent RAG retrieval should show what was retrieved and why.

Possible retrieval paths include local embeddings, a local server, or a cloud service. The choice should be visible, privacy-aware, and compatible with offline behavior. Source deletion must support re-indexing and removal of derived index data. Storage footprint should be bounded, with indexes treated as regenerable artifacts rather than a second silent authority.

A source quote or citation is not the same thing as an AI inference. The workflow must avoid fabricated citations, distinguish quotation from interpretation, and never silently insert retrieved or generated content into the manuscript. Authors should explicitly review and insert material.

Context

This Discussion connects product exploration with the storage/recovery context of #482, browser-local AI context of #483, Capture Inbox context of #484, and local model lifecycle/Eco-Mode context of #556. Those Issues remain authoritative for their implementation scopes; no new Issue crosslinks are added here.

Role perspectives

Role-perspective note: The viewpoints below are maintainer-curated, AI-assisted design lenses. They are not separate community members, votes, user testimony, or evidence of consensus.

• 🔎 Research-heavy nonfiction author: reliable provenance, citations, excerpts, and repeatable source review.
• ✍️ Fiction/worldbuilding author: useful background and continuity without contaminating canon or manuscript voice.
• 🔐 Privacy/local-first: offline operation, local indexes, explicit provider boundaries, and deletion that really removes derived data.
• 🤖 RAG/AI product: transparent retrieval, inspectable context, useful ranking, and clear regeneration behavior.
• 🧪 Provenance/reliability: distinguish source evidence from inference and fail closed when citations cannot be verified.

Questions

• What should move from Capture Inbox into a research library, and what should remain an ephemeral note?
• Which provenance and citation details are essential during writing?
• How transparent should retrieval chunks, ranking, and model-generated inferences be?
• When are local embeddings, a local server, or cloud retrieval acceptable?
• What should deletion and re-indexing prove?
• What storage footprint and offline behavior would make the workflow trustworthy?

Relationship to implementation

This Discussion is research and workflow design. It does not authorize silent manuscript insertion, fabricated citations, a second source-of-truth, or changes to #482, #483, #484, or #556.