How Redil works.

The path of a request

1. The request arrives through an API at Redil's infrastructure, inside the EU.

2. The router, an open decision model, classifies it and picks the base model and the sources that handle it best.

3. Retrieval fetches the relevant passages from your private index and from legislation.

4. The base model answers with your adapter loaded, which sets format, citations and terminology.

5. The answer comes back with its sources and the name of the model that produced it.

Models

Only open-weight models under permissive licences (Apache-2.0, MIT). We re-check the licence at every release.

Models run only on our infrastructure. No data reaches the people who made them, and you can restrict which models you accept.

Adapters

A LoRA adapter is a few megabytes of weights trained on your data. It is loaded onto a shared base model when a request comes in.

Because the base is shared, thousands of adapters can live on the same GPUs. That is the engine we are building: scheduling, caching and isolation between customers.

Retrieval over your documents

Every customer has its own index, separate from the rest. It combines keyword and semantic search and re-ranks the results before they reach the model.

Your documents supply the knowledge and the adapter sets the behaviour. Change a document and the answer changes, with nothing retrained.

Security and data

EU-owned infrastructure. No company subject to the US CLOUD Act anywhere in the chain.

Your data, indexes and adapters are yours, isolated per customer, and train no other model.

Every deployment ships with its documentation for the GDPR and the AI Act. Spain's ENS certification is on the first-year roadmap.

API

The API will follow the chat format most tools already use, so that moving to Redil means changing a URL and a key.

It is not available yet. This is what a call will look like:

curl https://api.redil.ai/v1/chat/completions \
  -H "Authorization: Bearer $REDIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "auto",
    "messages": [{"role": "user", "content": "What is the deadline to appeal?"}]
  }'