Installation

Install the Ovrin core and only the provider modules you need. The core has zero dependencies and no cgo, so it cross-compiles and builds static.

Ovrin is a Go library. The core module has zero external dependencies and uses no cgo, so it
cross-compiles and builds static. Provider-specific code lives in separate modules, so you pull in
only the vendors you actually use.

Before you start: no release is tagged yet

Ovrin is pre-v1 and no version has been tagged. The go get commands you would expect will not
resolve yet, and the API will change before v1.0. Until a release exists, build from a checkout.

git clone https://github.com/BAGOMBEKA-JOB-DEV/ovrin.git

Then point your module at that checkout with a replace directive:

// go.mod
require github.com/BAGOMBEKA-JOB-DEV/ovrin v0.0.0

replace github.com/BAGOMBEKA-JOB-DEV/ovrin => ../ovrin

This is the same arrangement the provider modules use among themselves while the repository is
untagged.

The modules

The core gives you Client, Extract, the three provider seams, and the whole pipeline. Everything
else is optional and lives in its own module with its own dependencies.

Module What it gives you
ovrin The core library. Zero dependencies, no cgo.
ovrin/model/skyl The Model seam over OpenAI, Gemini, Ollama, and OpenAI-compatible hosts.
ovrin/ocr/google OCR via Google Cloud Vision. Standard library only.
ovrin/ocr/tesseract Local, offline OCR. Tesseract compiled to WebAssembly, so still no cgo.
ovrin/ocr/azure OCR via Azure AI Document Intelligence. Reports table structure.
ovrin/ocr/textract OCR via Amazon Textract. Standard library only, SigV4 included.
ovrin/render/pdfium The Renderer seam. PDFium under WebAssembly, for rasterising PDF pages.
ovrin/otel Turns Ovrin hooks into OpenTelemetry spans and metrics.

Once releases are tagged, each module is versioned independently on a path-prefixed tag, and each
declares its own Go floor.

Go version

The core module declares go 1.22 as its language floor. Because Ovrin hands untrusted document
bytes to archive/zip, encoding/xml, compress/flate, and image/*, build with a current patch
release rather than the oldest toolchain that compiles.

What you actually need

Most extractions need only the core plus a model adapter. You need a renderer only when you are
running OCR locally against a scanned PDF: a cloud OCR provider that accepts a PDF rasterises it
server-side, and image inputs never need rasterising at all.

Next