Skip to content

Changelog

Changelog

What landed in this wiki, newest first.

2026-10-12

  • How DPT-3 works: evidence-graded reconstruction with documented / implied / unknown chips, version timeline, and black-box probes to run with an API key.
  • Korean slides pipeline: six stages, five-second budget table, Korean prompts, resolution math; reference code in pipeline/korean_slides/ with a dry-run mode and a verify.decide() policy left for the reader.
  • Model candidates: Qwen3-VL, Qwen3.5, Qwen3.8 and the 1B parsers (PaddleOCR-VL-1.6, OvisOCR2, HPD-Parsing, dots.ocr, DeepSeek-OCR) with code and published results, and the recommended combination.
  • Embeddings and retrieval: two-track design, the KoVRE Korean benchmark table, chunking from the ADE-shaped output, encoding and fusion code.
  • Hardware and serving: what fits on 2x/4x A100 40/80 GB, vLLM commands, Qwen3.5/3.8 caveats on A100, vLLM Semantic Router as the serving-side decision layer, TPU as a hypothesis, M1 Max for development.
  • Evaluation recipe: eval-set buckets, ground-truth protocol, CER/order/TEDS/rubric metrics, pipeline/eval/compare.py, pipeline/eval/run_ade.py, timing harness.
  • Research: two-stage vs end-to-end, agentic repair loops, slide understanding, benchmark blind spots; Korean benchmarks, models and text details.
  • What LandingAI built, The API contract, What REA revealed.
  • Site scaffold: MkDocs Material themed after docs.landing.ai, deployed by GitHub Actions.