You decide which files get indexed

Your folders often hold more than what you want AI to read: thousands of screenshots, gigabytes of screen recordings, dependency packages, and private material you would rather keep away from agents. You can now filter by file type, and use exclusion rules to skip specific files and subfolders. On top of that, the official models get a full upgrade, and remote MCP now hides more than 200 kinds of sensitive information automatically.

1. Indexing supports file type filters and exclusion rules

Under Index → Rules, there are two ways to control what gets indexed:

• File type switches: PDF, Word, Markdown and plain text, web pages, ebooks, spreadsheets, presentations, images, audio and video can each be turned off. A type that is off is no longer scanned, and files already indexed are removed from the index. If you only want AI to read documents, for example, turn off Images and no time goes into recognizing the text in screenshots.
• Exclusion rules: filter with wildcards, such as node_modules/, *.log or private/. A rule can apply to all folders or to just one. No need to learn the syntax: right-click a file or folder in the library and choose "Add exclusion rule…".

Before you save a rule, you are shown how many files it will exclude and exactly which ones. Exclusion only affects the index; the files on disk are never touched. Removal starts about a minute after you save.

2. Official model upgrade

The official models have been refreshed:

• For everyone: new DeepSeek V4.1 Flash, Qwen3.8 27B and GPT-6 Luna.
• For Pro: new Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol and Gemini 4 Argon; DeepSeek V4 Pro is updated to the 0813 version.

Before an official model goes live, we test it on real libraries to see how it handles searching, reading and citing, and we compare model providers one by one. Starting with this version, apart from DeepSeek, official models only use providers with zero data retention, to keep your data safer.

3. Stronger protection: remote MCP hides more than 200 kinds of sensitive information

Security is our lifeline. This version adds an important upgrade: when a cloud agent reads your library through remote MCP, sensitive information in your material is automatically replaced with a [REDACTED:type] marker, so it does not leak by accident. The types recognized so far include:

• API keys and tokens for more than 160 services, including OpenAI, Anthropic, AWS and GitHub
• Private keys, JWTs, and passwords inside connection strings
• Common password strings such as "password: xxx" written in notes, or PASSWORD=xxx
• Common ID numbers: bank card numbers (only the last four digits are kept), national ID numbers, passport and travel permit numbers
• Two-factor recovery codes and wallet recovery phrases

This greatly lowers the risk of leaking credentials, but it is not foolproof. For your most important secrets and private data, we recommend keeping them out of the index with the exclusion rules above.

4. Other improvements

• See the original in the app: document previews switch between the parsed view and the original file, showing PDF pages, images, audio and video directly.
• Double-click to open files: double-click a file in the library's file list to open it with the system's default app.
• Scanned PDFs keep their pictures: pictures and charts in scanned PDFs are now cropped, saved and linked from the extracted text.

See the changelog for the full list.