[FEATURE] Need a RAG toggle in Projects

Resolved 💬 3 comments Opened Apr 12, 2026 by chopmix Closed Apr 15, 2026

Preflight Checklist

  • [x] I have searched existing requests and this feature hasn't been requested yet
  • [x] This is a single feature request (not multiple features)

Problem Statement

When I upload files to a Claude Project's knowledge base, if the total file size fits within the 200K context window, all files load in full into every conversation. This eats my working context — leaving less room for deep, multi-turn work.

The only way to trigger RAG (retrieval-on-demand) is to upload enough files to exceed the context window threshold. This forces power users to artificially bloat their knowledge base just to get selective retrieval behavior.

My workflow: I run 11 Projects across a consulting practice. Some need a small but focused reference library (5-10 pages of frameworks, vocabulary maps, pricing docs). I want those files searchable on demand, not loaded in full every time. Right now I have to choose between keeping files tiny (losing reference depth) or overstuffing to trigger RAG (wasting storage and degrading retrieval quality with filler).

Proposed Solution

Add a toggle in Project settings: "Use RAG retrieval" (on/off).

  • OFF (default): Current behavior. Files load in full into context.
  • ON: Files are indexed and retrieved on demand via the project knowledge search tool, regardless of total file size.

This is a one-toggle UX change. The RAG infrastructure already exists — it just needs a manual override instead of only activating at a size threshold.

Alternative Solutions

Workaround today: Create compressed markdown "essentials" files that distill large reference docs into 1-2 pages, and only upload those. This works but requires manual maintenance every time a source document updates. A RAG toggle would eliminate this overhead entirely.

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

_No response_

Additional Context

_No response_

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗