Skip to content

Implementation status

This page describes the capability boundary users and contributors should rely on today.

Available

  • Local Tauri desktop shell with Svelte UI and Rust commands.
  • Password-protected encrypted vault with lock/unlock and persisted unlock backoff.
  • Optional passkey PRF registration and unlock where the platform supports it.
  • New supported-platform vaults can use an OS-vault-only first-run path, with passkey management and authenticated master-key rotation available in Settings → Security.
  • Settings → Security controls for optional Windows Credential Manager, macOS Keychain, or Linux Secret Service vault wrapping, with explicit automatic unlock control.
  • PDF import, SHA-256 duplicate detection, PDFium extraction, source-PDF retention, parsing, normalized rows, and audit diagnostics.
  • Patient and report organization, global search, comparison views, charts, CSV export, Library controls, and storage backup/restore flows.
  • Cross-platform CI and a release matrix covering the supported desktop architectures.

Available with conditions

  • Tesseract OCR is compiled into distributed builds. A native executable can be bundled with TESSERACT_BUNDLE_DIR; Settings can download the eng/por language data, while an installed executable remains a supported fallback.
  • Ingestion stages are independently controlled in Settings: Tier 1 PDFium extraction, Tier 2 Tesseract OCR fallback, and Tier 3 hybrid OCR + LLM escalation. Tier 3 can load Phi-4 on demand and persist a Tier 3 report only when repaired output was actually used; missing or failed optional models fall back to the successful lower tier with diagnostics.
  • PDFium may need a target-matched manual library when the build-time download is unavailable.
  • Signing depends on platform credentials and repository secrets; default artifacts can be unsigned.

In progress or experimental

  • Phi-4 uses the bundled llama.cpp integration for local text repair when a compatible GGUF model is configured and loaded. olmOCR-2 model management and readiness reporting are present, but its safetensors vision inference runtime is not yet connected to ingestion.
  • Production confidence still depends on expanding synthetic and sanitized PDF fixture coverage across real-world report layouts.

For the full design history and acceptance criteria, read the original implementation plan.