EVIE 4.5B vs EVIE 8B

Both EVIE checkpoints use multi-vector visual-document retrieval and late-interaction scoring. They are appropriate when questions depend on a page’s visual structure, tables or figures. The 4.5B release documents shorter Matryoshka dimensions and token-compression options; the 8B release uses larger token vectors. These differences affect the index and scoring workload as well as the encoder.

Use the same page images and labelled questions to build a baseline for each checkpoint. Include table-coordinate questions, chart relationships and pages with near-identical headings. Score whether the page containing the needed evidence is retrieved, then inspect ranking mistakes. Treat compression and dimension reduction as separate experimental settings rather than assuming they preserve every query’s result.

A conventional single-vector index does not reproduce the documented retrieval method. Measure storage and latency with a backend that supports the model’s representation, and preserve page rendering and scoring precision across comparisons.