You've been working on a project with Claude. You've uploaded your documents, your notes, your reference files. Claude confirmed they were indexed and ready. Then halfway through a session it starts giving you answers that have nothing to do with what's in those files — or worse, answers based on files you deleted weeks ago.
This is one of the most reported bugs with Claude Projects right now. And it has a name: RAG switching. Understanding what it is changes how you work with Claude permanently.
What is RAG switching and why does it happen?
When you upload files to a Claude Project, Claude starts by reading them directly — the whole thing, right into its active memory. This works brilliantly for small file collections. Claude has everything it needs right in front of it.
But around 13 files, something changes. Claude quietly switches from reading your files directly to searching them — a system called RAG, which stands for Retrieval-Augmented Generation. Instead of having everything in memory, Claude now has to search for relevant pieces when you ask a question.
The problem is that this search is broken. Users report files showing as "fully indexed" when they haven't been indexed at all. Searches return nothing. And deleting and re-uploading files often doesn't fix it — Claude continues referencing the old deleted content from a cache it never cleared properly.
The result is Claude answering confidently from files that no longer exist, or ignoring files that do. It's not lying. It's searching a broken index and reporting what it finds.
Claude is the engine. You're the driver.
Here's the way to think about your relationship with Claude on any project that involves files.
Claude is the engine. Powerful, capable, does the heavy work. But the engine doesn't decide where you go.
You are the driver. You provide the inputs — the direction, the instructions, the decisions.
Your file system is the gear shift. It determines what mode you're operating in. And just like driving, you still have to put the car in the right gear depending on what's needed. Reverse for going back. Drive for moving forward. Neutral when you need to pause.
Most people hand Claude the keys and expect it to drive itself. It can't. It needs you in the seat, managing the gear shift, deciding what files are active, what gets saved, and what gets left behind.
This isn't a criticism of Claude. It's just how the system works right now. And if you know that going in, you can build around it.
The folder system that solves this
If there's a known risk that files might get lost, corrupted, or ignored — the answer isn't to hope it doesn't happen. The answer is to manage your files properly on your end so that when it does happen, you lose nothing.
This is the folder structure we recommend to every client working on ongoing projects with Claude:
Your project folder structure
The habit that protects everything
The single most important rule: always work from the last saved file, never from memory. Before every Claude session, check your Current folder, confirm the file versions are right, and upload fresh. Don't assume Claude remembered. Don't assume the Project index is accurate. Treat every session like a fresh start with fresh files.
Before every session
Open your Current folder. Check the files are the latest versions. Upload them fresh to Claude — don't rely on what's already in the Project.
After every session
Save any files Claude produced or modified into your Outputs folder immediately. If Claude edited an existing file, move the old version to Archive with today's date before replacing it.
Never delete — always archive
The temptation is to delete old versions to keep things tidy. Don't. Move them to Archive. If Claude's cache ever references an old file, you'll want it available to compare.
Keep your Project file count under 10
Stay well below the 13-file threshold where RAG switching kicks in. If you have more files than that, split them across separate Projects by topic rather than piling everything into one.
If Claude references a deleted file
Start a brand new Project. Don't try to fix the broken index — it won't work. A fresh Project with freshly uploaded files from your Current folder takes five minutes and solves it completely.
This is a team effort
The people who get the most out of Claude on complex projects aren't the ones who expect it to manage everything. They're the ones who treat it as a partnership. Claude does the thinking and the writing and the building. You manage the file system, the versions, and the context. Neither side can do the other's job.
The RAG switching bug will get fixed eventually — Anthropic is aware of it and actively working on it. But while it exists, the answer isn't frustration. The answer is a file system that makes the bug irrelevant. You manage the gear shift properly and the engine keeps running no matter what.
The most resilient Claude workflows we've seen are built by people who assume Claude might forget something and plan accordingly — not people who assume it will always remember. Build the habit before you need it.
If your current project is already in a broken state — Claude referencing wrong files, ignoring uploads, or working from deleted content — a session with us can help you rebuild the file structure cleanly and get back on track without losing your work.