BALISE

The problem

Finding a regulatory answer takes time, and an assistant that answers without citing its source is worse than useless: credible without being verifiable. In a regulated field, an untraceable answer is worth nothing.

How it works

The corpus is chunked, indexed and queried by vector search. The model answers only from retrieved excerpts, and every answer carries its references. Domain knowledge stays out of the core: a module ships separately, leaving no trace of its field in the engine.

Core capabilities

• Cited answers, tied to the articles retrieved from the corpus
• Deterministic calculations kept out of the language model, in verifiable code
• Interchangeable domain modules, shipped separately from the core
• Local execution by default; a cloud model activates only by decision
• Versioned validation sets, including an adversarial set written to make the engine fail

Stated limits

• Human validation is mandatory on answers as well as calculations
• No authentication: designed for a local machine, not to be exposed on a network
• Quality is proven on the versioned sets, not beyond: no pilot on real files
• A regulatory assistant replaces neither professional advice nor reading the standard

Technologies

• Python, Pydantic data models, command line interface
• Qdrant vector database, embeddings computed locally
• Ollama for the local stack, interchangeable model interface
• Modular architecture, boundary held by golden tests