Descriptions:
Fahd Mirza introduces Tencent’s EVIE (EV Preview), a 4.5 billion parameter visual document retriever that sidesteps OCR entirely by treating document pages as raw images. The model, built on a ColBERT/ColPali architecture with a shared backbone, encodes both text queries and document image patches into compact 128-dimensional token vectors — keeping one million pages indexed under 180 GB while achieving top rankings on the ViDoRe v3 and ViDoRe v1+v2 retrieval benchmarks, beating significantly larger models.
The video walks through a full local installation on Ubuntu using a UV environment and the Golde Pali engine, running on an NVIDIA RTX A6000. Mirza demos the model against an invoice image and a dense academic reference list, showing relevance scores of 15.19 and 14.12 on two questions — and clarifying the key architectural point: EVIE finds the right page, it does not extract the answer, so it pairs with a reader model in a two-stage pipeline.
The architecture explanation is notably accessible: the shared backbone processes text on the left and image patches on the right through the same weights, producing per-token and per-patch vector crowds rather than a single compressed embedding — a design that preserves fine-grained layout information that OCR destroys. VRAM consumption comes in under 10 GB at load, making local deployment practical for developers working with financial filings, scanned forms, or multilingual technical documents.
📺 Source: Fahd Mirza · Published August 24, 2026
🏷️ Format: Tutorial Demo
