Self-hosted latency benchmark request: tree generation stage is very slow on scanned PDFs

[Taooo-habitus]

I'm trying to benchmark the latency of self-hosted PageIndex, focusing specifically on the tree generation stage.

I'm using a local fork that incorporates concurrent request handling, but I would like to understand the baseline performance of the upstream self-hosted version for comparison, particularly regarding tree generation.

Environment

• Self-hosted deployment with concurrent LLM calls
• OCR input: scanned/image-based PDF
• Document length: 128 pages
• Model backend: Snowflake Cortex (Claude Sonnet 4.5)

Observed Latency

• Tree generation with concurrent requests: approximately 125–140 seconds (with concurrent requests set to 128)
• Tree generation without concurrent requests (original implementation): approximately 400 seconds

I have also tested the same document through the official chat page, which completed in around 60 seconds end-to-end. I understand that the official framework may have implemented many more optimizations, but I'm still curious whether the stats above are expected.