PS ID: ALG-AI-02 | Domain: AI / ML
A browser-native document investigation platform. Upload PDFs, DOCX, CSV, HTML, text, and images. Ask natural-language questions. Get cited answers, conflict detection, and calibrated uncertainty — without any API keys, servers, or cost.
- Ingests multiple document formats (PDF, DOCX, CSV, HTML, TXT/MD, images via OCR)
- Indexes them locally using real sentence embeddings
- Answers natural-language questions with source citations (file · heading · page)
- Detects conflicts between documents (two independent layers: NLI + rule-based)
- Communicates uncertainty with a 4-level confidence gate instead of guessing
| Feature | Implementation |
|---|---|
| Multi-format ingestion | pdfjs-dist, mammoth, papaparse, Tesseract.js, DOMParser |
| Local embeddings | Xenova/all-MiniLM-L6-v2 (22 MB, ONNX, WASM/WebGPU) |
| Persistent index | IndexedDB (survives page reloads) |
| Semantic retrieval | Cosine similarity over 384-d vectors |
| Conflict detection | Xenova/nli-deberta-v3-xsmall (NLI) + numeric rule-based fallback |
| Uncertainty handling | 4-level confidence: HIGH / MEDIUM / LOW / INSUFFICIENT |
| Source grounding | file · heading · page · similarity score |
| Privacy | 100% client-side — documents never leave the browser |
| Cost | $0 — no API keys, no per-query charges |
git clone https://github.com/YOUR-USERNAME/localmind.git
cd localmind
npm install
npm run dev
# Open http://localhost:3000