DESelf-host

Models and hardware

Docuvate is CPU-first: OCR, embeddings, and the default chat run without a GPU. Heavier models are optional and documented in settings.

OCR and extraction

  • Default: PaddleOCR PP-OCRv4 (Latin script, including German) in the worker.
  • Text-layer PDFs skip OCR; Tesseract or Docling are optional.
  • Heuristic fields (amount, date, sender) from extracted text.

Embeddings and search

Document embeddings via the worker (paraphrase-multilingual-MiniLM-L12-v2). Label suggestions use cosine similarity, not an LLM.

Full-text search with optional typo tolerance (pg_trgm) and field filters; semantic components per repository configuration.

Document chat

Variables: DOCUMENT_CHAT_PROVIDER, OLLAMA_URL, OLLAMA_MODEL, WORKER_URL. Details: docs/ai-models.md in the repository.

  • Compose default: RAG over extracted text plus Ollama (for example qwen2.5:1.5b on CPU).
  • Donut/GPU paths are optional and enabled in the UI only with suitable hardware.
  • No sending your files to cloud LLM vendors on the product path.