Where your data goes
Every project runs in one of two places. On our servers: your data comes to hardware we own and operate, and is processed there. On your hardware: for sensitive work, we install and run the models on your organization’s own machines, so your data never leaves your systems.
What we never do
- Send your data to commercial AI companies.
- Rent AI by the request from a vendor whose product can change underneath you.
Open-weight models
We run open-weight models we can inspect and control. Nothing depends on a vendor’s black box or pricing. Models we run include Llama, Qwen, Mistral, Gemma and DeepSeek, served on our own hardware through Ollama.
Capable, but not frontier models
The models we run are capable: they read contracts, match records, draft documents, provide legal analysis and cite the cases behind it. But they aren’t “frontier” models, the largest and newest systems that big AI companies train at enormous cost and run in their own data centers. Safety testing of some frontier systems has reported troubling behavior in test scenarios, such as trying to avoid being shut down or to copy themselves to other servers. The models we use are smaller, and do the specific jobs we give them, inside our own systems.
No new data centers
The work runs on hardware that already exists. It doesn’t add to the demand that is driving new data center construction.
Retrieval comes first: how we build legal RAG
Our cited answers use retrieval-augmented generation, or RAG. Instead of answering from memory, the system first retrieves the passages that bear on a question (contract articles, board decisions, statutes), then the model writes its answer from those passages and cites them. Legal RAG applies this to legal text, where an answer is only as good as the source it can be checked against.
Two recent studies show that the retrieval step matters as much as the model, if not more. LegalBench-RAG (Pipitone and Houir Alami, 2024) built the first benchmark for the retrieval step of legal RAG. It found that how documents are split into passages changes the results, and that general-purpose tools did poorly on specialized legal text. Legal RAG Bench (Butler and Butler, 2026) tested complete legal RAG systems. It found that the quality of retrieval, more than the choice of language model, drove correct answers, and that many errors blamed on a model making things up began with a retrieval failure.
So we build retrieval with the same care as everything else. Texts are split meticulously along their own structure, by article, section and subsection, so each passage is a complete provision. Passages are indexed with embedding models chosen for how well they find legal language. And each job uses the local model best suited to it.
Questions your IT or legal team may ask
- Which models do you run? Open models such as Llama, Qwen, Mistral, Gemma and DeepSeek, run locally through Ollama.
