Skip to content

vera-app

vera-app is the VERA desktop product: an Electron and React interface backed by a local Python sidecar. It composes vera-doc for retrieval, vera-ingest / vera-ingest-pymupdf for PDF conversion (Markdown ingest is bundled in vera-ingest), and vera-embed-openai for hosted OpenAI embeddings; it does not use the CLI as its backend. vera-ingest-docling is an optional CLI extra, not bundled in the installer.

See the product overview for the intended workflow and audience.

The Python package root intentionally exports no public API. Sidecar, LLM provider, mode, and cancellation modules are implementation details and are therefore documented in the architecture guide rather than generated as public API reference.

Install the Windows app

Download VERA.Setup.<version>.exe from the latest GitHub Release.

First workflow

  1. Open Convert and convert selected PDFs or Markdown files, or a directory, or right-click a folder and choose Convert…. Expand Advanced pipeline options for schema-driven settings from describe_ingest_pipelines / PipelineConfigForm. Convert lists PyMuPDF and the bundled Markdown pipeline. Convert embedding presets are hashing (default), Local semantic (MiniLM), and OpenAI text-embedding-3-*. MiniLM is ONNX Runtime under the same sentence-transformers/all-MiniLM-L6-v2 identity, and the installer vendors a VERA-exported, SHA256-pinned graph. Save OPENAI_API_KEY under File > Settings → Embeddings. Archives converted with OpenAI are not portable for semantic search. Right-click a .vera archive and choose Reconvert… to replace it with a different ingest pipeline or embedding model. Reconvert writes to the clicked archive even when its name differs from the source file.
  2. Use File > Open Folder to activate a document library.
  3. Open Search for fully local hybrid retrieval.
  4. To use Ask, configure a provider under File > Settings → LLM Providers.
  5. Optional: save a Hugging Face token under File > Settings → Hugging Face (or set HF_TOKEN) for Hub model downloads used by some converters and embedders. File > Open convert log... (or Convert Open log / Settings → Diagnostics) opens userData/logs/sidecar.log for timed convert steps.
  6. Select a citation in an answer to inspect the highlighted source passage. The PDF viewer uses Mozilla-style dark chrome with a page thumbnail rail, rotate counterclockwise, download, and print. Citation highlights stay aligned after rotate because they turn with the page. Ask answers render GitHub-flavored Markdown and LaTeX (KaTeX).

Search and conversion do not require a model-provider account unless you choose a hosted embedder such as OpenAI. A Chat provider is only required for generated Ask responses.

Source-run and packaged conversions use one sidecar interpreter with PyMuPDF, hashing, ONNX MiniLM, and OpenAI embeddings. Extra ingest and embedding plugins are pip packages in that same environment (python -m pip install or python -m pip install -e <clone>). See Creating an ingest pipeline plugin, Creating an embedding provider, and Desktop app architecture.

Documentation

Developer entry point

npm run app:install
npm run app:dev

The supported packaged target is currently Windows. Use the CLI --pipeline-option flags (or Convert-view pipeline settings) for provider-owned chunking and OCR controls. Packaged and app:dev Convert both report pymupdf, plus hashing and MiniLM embedders. app:dev vendors MiniLM into packages/vera-app/build/minilm before launch; packaged builds vendor a VERA-exported ONNX graph. The sidecar does not import Torch. Docling is not listed; use vera[docling] from the CLI.