Improvement vectors for DeepSeek-OCR

1. Compression ratio is fixed/manual, not adaptive
Current: user picks resolution mode (Tiny/Small/Base/Large) upfront — a static compression-vs-fidelity tradeoff.
Improvement: content-aware dynamic token allocation — dense text regions get more vision tokens, whitespace/background gets fewer. Right now compression is uniform across the image regardless of information density.
Trade-off: adds a routing/scoring overhead before encoding — costs latency to save tokens. Only worth it if token cost >> routing cost.

2. Lossy compression has no confidence signal
Current: decoder just outputs text, no indication of which regions were reconstructed with low confidence (i.e., which compressed tokens were "guessed" vs. "read").
Improvement: expose per-token/per-region confidence scores on the decode side, so downstream RAG pipelines know where to fall back to re-OCR at higher resolution or flag for human review.
This is the single highest-leverage fix if you're actually deploying this — silent hallucination on compressed regions is the real production risk, not raw accuracy.

3. No native handling of multi-page document coherenceCurrent design is page/image-level. Cross-page context (headers/footers, running tables, references) isn't modeled.
Improvement: a lightweight cross-page attention or memory mechanism so compressed tokens from page N can attend to page N-1's compressed tokens — genuinely useful for the "long-context via images" framing since multi-page docs are the actual use case.