codeflash-agent/.codeflash/unstructured/core-product/status.md

core-product Status

Last updated: 2026-04-10

Current state

Stacked PR #1500 updated with cumulative progression (proper benchmarking: 5 rounds + 1 warmup, median reported, <0.4% stddev). Cumulative: 14.6% latency on 10p-scan, 13.3% on 16p-mixed, 2.1 GB memory savings. Next: optimization #4 (direct numpy-to-BMP for tesseract).

Target repo

~/Desktop/work/unstructured_org/core-product on branch main (PR branch: perf/cpu-aware-serial-ocr)

PRs

PR	Branch	Status	Description
#1503	perf/bmp-render-format	Draft	Render PDF pages as BMP instead of PNG in pdfium pool
#1502	perf/cpu-aware-serial-ocr	Draft	Cap OCR workers to available CPUs (serial mode on 1-CPU pods)
#1500	codeflash-agent	Draft	Stacked optimizations + benchmark infra (cumulative progression)
#1481	perf/elements-intersect-vertically	Merged	Reduce attribute lookups
#1464	replace-lazyproperty-with-cached-property	Merged	Replace lazyproperty with functools.cached_property
#1448	mem/free-pil-before-table-extraction	Merged	Free page image before table OCR
#1441	mem/numpy-preprocessing-yolox	Merged	Resize-first preprocessing
#1400	async-join-responses	Merged	Fix blocking event loop in CSV merge

Optimization queue

1. CPU-aware serial OCR — PR #1502 open (draft), benchmarked. Rebase after #1501 merges.
2. Early memory release — skipped, codebase already well-optimized (context managers, per-page cleanup)
3. BMP render format — PR #1503 open (draft), benchmarked. 14.9% latency improvement on 10p-scan.
4. Direct numpy-to-BMP for tesseract — encode from numpy without PIL round-trip
5. Skip remaining PIL↔numpy conversions in OCR path

Dependencies

• PR #1501 (segfault fix, patched_convert_pdf_to_image refactor) must merge before #1502 rebase. Different functions, clean rebase expected.

VM

• IP: 40.65.91.158
• Size: Standard_E4s_v5
• RG: core-product-BENCH-RG
• State: Running (verified 2026-04-10)
• Git auth: HTTPS with embedded token (set previously). Use ssh -A for agent forwarding if token expires.
• Note: uv is at ~/.local/bin/uv — needs export PATH=$HOME/.local/bin:$PATH in non-login shells.
• Note: pytest-benchmark installed in .venv (not in lockfile).

Next steps

1. Implement "pass file path to tesseract" optimization (skip PIL→numpy→PIL→temp-file round-trip)
2. Benchmark on VM, open draft PR
3. Rebase #1502 once #1501 merges

Notes

• memray tree opens a TUI — do not run directly over SSH. Use memray stats, memray summary, or memray flamegraph --output file.html instead.
• memray peak is 1.0 GB (10p scan, serial path). 10 GB total allocated = heavy PIL churn per page, not accumulation.
• Benchmarking: use pedantic(rounds=5, warmup_rounds=1) — warmup absorbs ONNX JIT + page cache. Observed stddev <0.4% of median. Guest CPU frequency controls are ineffective on Azure Hyper-V — use statistical methods (more rounds + median) instead of trying to pin frequency.
• Workflow: independent perf/<name> branch → open individual draft PR → cherry-pick to codeflash-agent → benchmark stacked progression → update #1500 body.