[BUG] Project files switch to RAG search at 2% capacity — threshold based on file count, not token size

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Claude Projects switch from direct context loading to RAG search mode (project_knowledge_search) at approximately 13 files, regardless of total token size. The UI displays “To save space in chats, Claude will look up specific information as needed” at just 2% of project capacity. This contradicts the official documentation which states: “RAG activates automatically when your project knowledge approaches the context window limit.”
My project contains 15 text files totaling approximately 73,000 tokens — 35% of the 200K context window. This is nowhere near the context window limit, yet the system forces RAG search for all files.
Steps to Reproduce:

1. Create a new project
2. Add 10 text files (any size, I tested with files ranging from 100 to 13,000 tokens)
3. Open a new chat — observe that files load directly into context. UI shows “1% of project capacity used” with no warning message
4. Add 3 more text files to the same project (total now 13 files, same approximate total token count of ~73K)
5. Open a new chat — UI now shows “2% of project capacity used” AND displays the warning: “To save space in chats, Claude will look up specific information as needed”
6. Claude now uses project_knowledge_search tool instead of reading files directly

Key observations from testing:
∙ At 10 files: direct context loading, no search, no warning
∙ At 12 files: direct context loading still worked
∙ At 13 files: RAG mode activates with warning message
∙ Removing the largest file (13,121 tokens) while keeping 14 smaller files: RAG still active — so the threshold is NOT based on individual file size
∙ Total token count remains ~73,000 in all configurations — only the file count changes the behavior
∙ The threshold appears to evaluate file count rather than total token size, which means a project with 13 files totaling 10,000 tokens would trigger RAG while a project with 9 files totaling 90,000 tokens would not
Expected Behavior:
Files totaling 73,000 tokens (35% of the 200K context window) should load directly into context, as documented: “RAG activates automatically when your project knowledge approaches the context window limit.” 35% is not approaching the limit. The system should evaluate total token size versus context window capacity, not file count.
Actual Behavior:
RAG activates at 13 files / 2% displayed capacity. All files are searched via project_knowledge_search instead of being loaded into context. The search retrieves partial fragments, misses cross-file connections, and causes the model to hallucinate details that contradict the actual file contents (wrong character names, invented locations, incorrect family relationships, fabricated data).
Evidence of regression:
When Claude Projects launched in June 2024, developer Simon Willison tested the system by uploading 63% of knowledge base capacity and reported that everything loaded directly into context. He noted that Claude “paste the whole thing into the context window.” In February 2026, the same system triggers search mode at 2%. The effective usable capacity has decreased dramatically.
Additionally, approximately one month ago, my same project with 17 files loaded all files directly without search. The threshold appears to have been changed silently without documentation updates.
Impact:
This bug makes Projects unusable for any sustained complex work requiring more than 12 reference files. In my case, a large creative writing project with interconnected character, economic, geographic, and political reference files. I cannot expand my files (adding character details, new locations, updated descriptions) because every additional token worsens RAG behavior. I am forced to write files in English instead of my native language to save tokens, and I manually count tokens using an external tokenizer because the UI provides no per-file token information. The RAG search produces worse results than direct loading — partial retrieval leads to hallucinations, and project instruction adherence degrades significantly when the model only sees file fragments instead of complete files.
Screenshots:
[IMG_4929] — 10 files, 1% capacity, no warning, direct loading
[IMG_4930] — 13 files, 2% capacity, warning message “To save space in chats, Claude will look up specific information as needed”

What Should Happen?

Claude Projects switch from direct context loading to RAG search mode (project_knowledge_search) at approximately 13 files, regardless of total token size. The UI displays “To save space in chats, Claude will look up specific information as needed” at just 2% of project capacity. This contradicts the official documentation which states: “RAG activates automatically when your project knowledge approaches the context window limit.”
My project contains 15 text files totaling approximately 73,000 tokens — 35% of the 200K context window. This is nowhere near the context window limit, yet the system forces RAG search for all files.
Expected Behavior:
Files totaling 73,000 tokens (35% of the 200K context window) should load directly into context, as documented: “RAG activates automatically when your project knowledge approaches the context window limit.” 35% is not approaching the limit. The system should evaluate total token size versus context window capacity, not file count.
Actual Behavior:
RAG activates at 13 files / 2% displayed capacity. All files are searched via project_knowledge_search instead of being loaded into context. The search retrieves partial fragments, misses cross-file connections, and causes the model to hallucinate details that contradict the actual file contents (wrong character names, invented locations, incorrect family relationships, fabricated data).
Impact:
This bug makes Projects unusable for any sustained complex work requiring more than 12 reference files. In my case, a large creative writing project with interconnected character, economic, geographic, and political reference files. I cannot expand my files (adding character details, new locations, updated descriptions) because every additional token worsens RAG behavior. I am forced to write files in English instead of my native language to save tokens, and I manually count tokens using an external tokenizer because the UI provides no per-file token information. The RAG search produces worse results than direct loading — partial retrieval leads to hallucinations, and project instruction adherence degrades significantly when the model only sees file fragments instead of complete files.
Screenshots:
[IMG_4929] — 10 files, 1% capacity, no warning, direct loading
[IMG_4930] — 13 files, 2% capacity, warning message “To save space in chats, Claude will look up specific information as needed”

<img width="645" height="1398" alt="Image" src="https://github.com/user-attachments/assets/7d0b29b2-9d10-4aee-ae0b-209f0f8813bf" />
<img width="645" height="1398" alt="Image" src="https://github.com/user-attachments/assets/8828c6af-5668-4d14-b7d2-89c806a1cbed" />

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce:

1. Create a new project
2. Add 10 text files (any size, I tested with files ranging from 100 to 13,000 tokens)
3. Open a new chat — observe that files load directly into context. UI shows “1% of project capacity used” with no warning message
4. Add 3 more text files to the same project (total now 13 files, same approximate total token count of ~73K)
5. Open a new chat — UI now shows “2% of project capacity used” AND displays the warning: “To save space in chats, Claude will look up specific information as needed”
6. Claude now uses project_knowledge_search tool instead of reading files directly

Key observations from testing:
∙ At 10 files: direct context loading, no search, no warning
∙ At 12 files: direct context loading still worked
∙ At 13 files: RAG mode activates with warning message
∙ Removing the largest file (13,121 tokens) while keeping 14 smaller files: RAG still active — so the threshold is NOT based on individual file size
∙ Total token count remains ~73,000 in all configurations — only the file count changes the behavior
∙ The threshold appears to evaluate file count rather than total token size, which means a project with 13 files totaling 10,000 tokens would trigger RAG while a project with 9 files totaling 90,000 tokens would not

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

February 12, 2026 — 15 files loaded directly without RAG

Claude Code Version

N/A — this bug is in the Claude iOS app / Claude.ai Projects interface, not Claude Code CLI

Platform

Other

Operating System

Other

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_