V

vision-llm

A robust CLI that turns recordings and PDFs into clean, structured Markdown notes end-to-end.

Audio — chunks with configurable length and overlap, transcribes via a dedicated Whisper STT endpoint or cascades to OpenAI-compatible multimodal and Native Ollama backends.

PDF — text PDFs pass through pdftotext straight to refinement; scanned PDFs are auto-detected by character count, rendered to page images, and read by a vision model (or tesseract via an explicit --ocr escape, never silently).

LLM refinement — dedupes overlapping boundaries and formats the raw transcript into clean notes.

Resumable content-addressed cache — keyed by input hash with per-chunk/page signatures; a changed parameter invalidates only what depends on it.

Config resolves CLI > env var > settings file > defaults, including persistent chunk sizes. Python 3.12, built with uv, 147 tests at 100% coverage. Architecture is protocol-based and documented via TIPs (docs/).

anti-slop vision gates: prompt hygiene + VLM judge rubric (SHIP/FIX/REGEN)