opencomplai-ai

The optional AI intent classification plugin for Opencomplai. It adds the --ai-intent flag to opencomplai scan, classifying how each AI callsite in your code is actually used — its decision autonomy, the subjects it acts on, and which EU AI Act risk tier and Annex III area it maps to.

All inference runs locally — models execute on your machine via llama.cpp, or via a deterministic code-signal matcher that needs no model weights at all. No code or prompts leave your environment unless you explicitly opt into the saas backend.

Prerequisites

opencomplai-ai is a plugin. Install the core engine first:

pip install opencomplai-core   # or the opencomplai / opencomplai-cli suite

Install

# Base install — only the deterministic codebert-onnx matcher, no download
pip install opencomplai-ai

# Deep install — required for the default model (qwen2.5-coder-1.5b) and
# every other generative GGUF model
pip install "opencomplai-ai[deep]"

The base install alone can only run codebert-onnx. opencomplai scan --ai-intent resolves qwen2.5-coder-1.5b by default, which needs [deep] — without it the scan fails fast with an actionable message instead of downloading the ~1 GB model first and only then discovering it can't be run.

Usage

Once installed alongside the CLI, the --ai-intent flag becomes available on the scan command:

opencomplai scan --ai-intent

By default only callsites in files with lexical findings are annotated (fast). To analyze every callsite in the repository:

opencomplai scan --ai-intent --ai-deep

Useful flags:

Supported models

The default model is qwen2.5-coder-1.5b and requires the [deep] extra. GGUF models are downloaded from the Hugging Face Hub on first use and cached locally under ~/.cache/opencomplai/models/.

codebert-onnx is a deterministic Annex III / prohibited-practice / limited-risk code-signal matcher — no model weights, no download, runs on the base install. It trades recall for speed and zero setup; it is not the default.

opencomplai scan --ai-intent                              # default: qwen2.5-coder-1.5b, needs [deep]
opencomplai scan --ai-intent --ai-model codebert-onnx     # no download, no [deep] extra

Model download flow

On first use of a GGUF model, the plugin prompts before downloading and shows a progress bar; the download is refused up front (no prompt, no partial download) if [deep] isn't installed. codebert-onnx needs none of this in normal use — it does deterministic code-signal matching with no model artifact to fetch.

An explicit prefetch/export of codebert-onnx (e.g. via opencomplai ai configure) is still available for callers that want the artifact anyway: CodeBERT has no prebuilt ONNX build on the Hub, so that path exports the official PyTorch checkpoint to ONNX on first run. It needs the separate [onnx] extra (optimum[onnxruntime]) and is unrelated to --ai-intent classification.

Optional extras

Configuration

Documentation

Full AI-intent guide and the model reference at docs.opencomplai.com.

License

AGPL-3.0-only. See LICENSE.