StackDrop

Your Downloads folder is a graveyard. You remember a sentence from a document, but not its filename or where you saved it.

StackDrop searches the contents of your local files, then summarizes only the document you choose. Your entire library never leaves your computer.

What it does

StackDrop is a local-first Windows desktop app that indexes and searches documents by filename, folder path, and extracted content.

Users can:

• Search across TXT, PDF, DOCX, and DOC files
• Find documents based on words inside them
• View search snippets, file metadata, and parsing diagnostics
• Automatically re-index files when they change
• Open any document and generate a structured summary in one click

Indexing and search remain completely local. Summaries are optional and use the user's own OpenAI API key. Only the selected document's prepared text is sent to OpenAI.

How we built it

StackDrop uses:

• Tauri and Rust for the native Windows application
• React and TypeScript for the interface
• SQLite FTS5 for fast local full-text search
• Local PDF, DOCX, DOC, and OCR extraction tools
• GPT-5.6 through the OpenAI Responses API for document summaries
• Windows Credential Manager for secure BYOK API-key storage

For long documents, StackDrop deterministically samples the beginning, middle, and end instead of sending only the first section.

GPT-5.6 returns a validated structured summary containing an overview, key points, important details, dates, action items, and uncertainties.

We used GPT-5.6 in Codex to plan, implement, test, red-team, and document the summarization system.

Challenges we ran into

The hardest challenge was adding cloud-powered summaries without breaking StackDrop's local-first promise.

We had to ensure that:

• API keys never entered the frontend bundle or database
• Stored credentials were never returned to JavaScript
• Absolute file paths and unrelated documents were never transmitted
• Document text was treated as untrusted data to resist prompt injection
• Long documents stayed within strict request limits
• Malformed model responses could not break the interface
• The feature still worked inside a packaged Windows installer

We also kept the new feature isolated from the existing indexing and search pipeline, so a summary failure could never affect the user's local document library.

Accomplishments that we're proud of

We are proud that StackDrop is not just a prototype. It is a packaged, installable Windows application.

The final product includes:

• Fast local document indexing and full-text search
• Secure BYOK storage through Windows Credential Manager
• One-click document summaries
• A responsive and accessible summary side panel
• Structured and validated GPT-5.6 responses
• Explicit privacy boundaries
• Automatic folder watching and re-indexing
• Detailed parsing and indexing diagnostics

Our final verification included:

• 96 unit and integration tests
• 11 Playwright tests
• 27 Rust tests
• Successful frontend builds
• Security boundary checks
• Successful MSI and NSIS Windows packaging

What we learned

We learned that privacy does not have to mean removing useful AI features. The better approach is to make AI selective, transparent, and user-controlled.

We also learned that local retrieval and cloud intelligence can complement each other. StackDrop performs the broad, privacy-sensitive work locally, then uses GPT-5.6 only when the user deliberately selects one document.

Finally, we learned that strong AI integration requires more than calling a model. Secure credential handling, input boundaries, structured outputs, failure states, testing, and clear user consent are equally important.

What's next for StackDrop

Next, we plan to improve:

• Support for additional document formats
• Search ranking and typo tolerance
• Faster indexing for very large libraries
• Cancelable summary requests
• Optional locally stored summaries
• Stronger long-document summarization
• Installer signing to remove Windows SmartScreen warnings

Our long-term goal is to make StackDrop the fastest and most private way to understand the documents already living on your computer.

Built With

• app
• codex
• desktop
• rust
• tauri