Skip to content

[T-068] Image/figure analysis is slow end-to-end — measure the bottlenecks and decide how to speed it up #123

Description

@Ztein

Canonical ticket: docs/tickets/T-068-speed-up-image-analysis-throughput.md

Converting a figure-heavy document is dominated by vision-model calls — a single description call has been observed anywhere from ~0.5 s to 60–90 s depending on endpoint load, and documents carry dozens of figures (a chartbook can carry 60–90). Users experience figmark as "slow on the interesting documents." Note what already exists before rebuilding it: within-document descriptions already run in parallel on a ThreadPoolExecutor at concurrency.max_workers (default 8). The ticket asks to measure the bottlenecks first, then decide.

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions