The Almirah: a 7B local vision model and 3D cabinet that keeps my friend's family papers off the cloud
Description and use
The Almirah stores documents offline in a 3D cabinet, uses local Qwen2.5 and BGE‑M3 models, and answers queries without internet.
Problem
Uncle Rajesh relied on memory to track 35 years of family paperwork, leading to missed renewals and lost documents.
Outcome
Uncle Rajesh can now locate documents instantly, receive accurate renewal dates, and avoid cloud exposure, as evidenced by a 1.2‑second query response during handover.
Workflow
- Scan or upload documents
- Local vision model extracts fields (Qwen2.5‑VL‑7B)
- Structured schema normalization (Qwen2.5‑7B‑Instruct)
- Human‑in‑the‑loop verification highlights bounding boxes
- Store data in local SQLite with BGE‑M3 embeddings
- Query via Ask Almirah, grounded RAG returns answer with citation badges
Evidence
The Almirah is a 100% local, air‑gapped document sanctuary and interactive 3D cabinet built to run entirely on a laptop with zero internet connection.
In 1.2 seconds, running purely on the local laptop model, the answer popped up: "According to your Private Car Package Policy Schedule, the Expiry / Renewal Date is 2026-10-22."
The entire codebase is open‑source under the MIT license: https://github.com/Shreyansh00987/The-Almirah
- Record date: 2026-10-04T22:41:25.771Z