Most AI spreadsheet and BI tools share the same architectural trade-off: to query your data with natural language, you have to upload your entire CSV or Excel file to their backend.
For quick public datasets, that's fine. For financial reports, customer PII, or internal production logs, uploading raw rows to a third-party server is often a non-starter.
I built Lumina Tabula to solve this by keeping the data processing engine inside the browser while still offering AI-assisted querying and visualization.
How It Works
Instead of shipping files to a cloud database, Lumina Tabula uses a three-part client-side pipeline:
-
In-Browser DuckDB-WASM Engine: When you drop a
.csvor.xlsxfile into the app, it is ingested directly into an in-memory DuckDB instance running inside a dedicated Web Worker. Running it in a worker keeps the main UI thread responsive even when scanning hundreds of thousands of rows. - Metadata-Only AI Context: To generate SQL from natural language prompts without exposing your dataset, the AI layer only receives the schema (column names, inferred data types) and lightweight summary statistics. Your actual rows never leave the browser sandbox.
- Client-Side Execution & Visualization: Once the model returns a SQL query, DuckDB-WASM executes it locally against your in-memory table and renders the output into interactive charts and summary views.
The Tech Stack
- Frontend: React + TypeScript
- Query Engine: DuckDB-WASM (WebAssembly)
- Styling: Tailwind CSS
- Auth & Sync: Supabase (optional authentication and saved workspace metadata)
Lessons & Challenges So Far
- WASM Memory Limits: Browser tabs have strict memory ceilings depending on the OS and device. Managing memory allocation and cleaning up temporary DuckDB tables when switching files was critical to prevent tab crashes on larger spreadsheets.
- Schema Ambiguity: When the LLM can't see sample rows (to preserve privacy), edge cases in date formats or dirty string columns require careful prompt engineering and automated SQL error-recovery loops.
Try It Out
You can test it with your own files (or sample datasets) at luminatabula.com.
Since everything runs locally in your browser, I'd love feedback from other devs on:
- Query latency on larger CSV/XLSX files on your machine
- Browser memory limits you run into
- Ideas for supporting local LLM endpoints (like Ollama) for 100% offline air-gapped workflows
Let me know what you think in the comments!
Top comments (0)