Skip to content

ROADMAP_2026_Q3.md

Latest commit

 

History

History
executable file
·
61 lines (48 loc) · 2.67 KB

ROADMAP_2026_Q3.md

File metadata and controls

executable file
·
61 lines (48 loc) · 2.67 KB
Raw

Omni Medical Suite — Q3 2026 Upgrade Plan

Scope completed in this refactor

  1. Core OCR modules merged into src/ocr/ and omni_medical_suite/preprocessing/

    • src/ocr/rtl_utils.py
    • src/ocr/field_extractor.py
    • src/ocr/deduplication.py
    • omni_medical_suite/preprocessing/compare_raw_vs_printed.py

  2. Routing expanded with explicit fallback chains

    • Handwriting → Arabic-handwritten-OCR (Qwen) → QARI → TrOCR
    • Vocalized Arabic → QARI
    • Structured extraction / forms → Nougat
    • Printed Arabic fallback → EasyOCR / PaddleOCR / Tesseract

  3. Patient-safe deduplication

    • Weighted fields: patient name, patient ID, date, diagnosis, medications, template signature
    • Same-template / different-patient cases are now intentionally penalized

  4. Semantic search path

    • QdrantMedicalSearch uses Qdrant when configured
    • Falls back to local fuzzy retrieval when vector dependencies are unavailable

  5. New experimental review UI

    • app/advanced_review_app.py added as an experimental review app (Compare / Search / Review tabs)
    • Not production-ready: missing image upload, Jais proofreading, HF Dataset save, translation, dictionary update
    • app/gradio_full_hitl.py remains the official production app (944 lines, all 10 features intact)

  6. Packaging and repository hygiene

    • pyproject.toml keeps setuptools.build_meta and now includes Qdrant support
    • root requirements-dev.txt reduced to a compatibility wrapper
    • .gitattributes now tracks data/**, *.jsonl, and *.parquet with Git LFS

  7. Regression coverage

    • RTL fixes
    • Field extraction
    • Advanced engine routing
    • Same-template / different-patient dedup edge case

Next development milestones

Milestone A — Runtime integration

  • Wire WeightedMedicalDeduplicator into API / job processors
  • Persist search index metadata alongside OCR jobs
  • Add upload-and-index flow in the review app

Milestone B — Benchmarks

  • Add CER/WER batch benchmark fixtures for raw vs preprocessed text
  • Store experiment outputs under a dedicated benchmark artifact path instead of the repo root
  • Validate Qwen/QARI/Nougat routing on representative medical samples

Milestone C — Human review loop

  • Save physician corrections in a single durable schema
  • Link corrected text back into search and dedup pipelines
  • Surface confidence + routing decisions in the UI for auditability

Milestone D — Repository cleanup

  • Remove obsolete duplicate scripts once imports are verified safe
  • Continue collapsing legacy requirement files into pyproject.toml extras
  • Move large benchmark fixtures fully behind Git LFS