In legal research, accuracy and confidentiality are paramount. DeepLaw combines sector-specific artificial intelligence with a controlled processing pipeline to deliver reliable and verifiable results. Let’s examine the technical foundations that make this possible.

An open-source model, specialised for Swiss law

DeepLaw is built on a fine-tuned open-source language model trained specifically on Swiss legal materials, including federal and cantonal legislation, court rulings and legal commentaries. In practical terms, this means the system has been adapted to understand the terminology, hierarchy and reasoning patterns of Swiss law rather than relying on general web knowledge.

From the next release onwards, the model will be updated monthly to integrate newly published laws and court decisions. A planned automated system will further ensure that legislative amendments and new rulings are incorporated consistently and systematically.

Swiss hosting and data protection

For legal professionals, confidentiality is not optional. DeepLaw is therefore hosted and operated entirely in Switzerland, with all data processed and stored in a secure cloud managed by DeepCloud. No data leaves the country, and no external public cloud providers are involved.

User data is stored solely to enhance the user experience, for example to resume conversations or share analyses with colleagues. It is not used to retrain or improve the underlying model. For particularly sensitive matters, the “Secret Chat” feature enables analyses without storing conversation data at all. In this model, conversation data is not stored on the platform and is deleted once the session is closed.

By combining Swiss hosting with strict data separation and controlled data elaboration, DeepLaw aligns with the confidentiality requirements of the legal sector from the outset.

How DeepLaw works

From a technical perspective, DeepLaw operates as a system composed of several stages.

The process begins with a structured reformulation of the user’s query into a clear legal case description. The system then performs a legislative analysis based on targeted searches within legal sources, identifying relevant federal laws from fedlex.admin.ch and, where applicable, cantonal laws from lexfind.ch. Legal commentaries from onlinekommentar.ch are incorporated where relevant.

Subsequently, the platform retrieves and analyses, via entscheidsuche.ch, federal and cantonal rulings in place since 1954. In total, DeepLaw draws on more than 21,000 laws, over 775,000 rulings and hundreds of expert commentaries. The final phase synthesises these elements into a structured conclusion outlining their legal implications.

Unlike generalist systems (such as ChatGPT, Copilot or other similar tools), which generate answers directly from statistical language patterns and draw on a wide variety of web sources, including blogs, websites and other online content that is not always verified, DeepLaw reasons iteratively on the basis of validated legal sources before producing a response.

Multi-step reasoning with full traceability

DeepLaw is designed for demanding legal workflows. It processes long documents progressively, enabling detailed analysis even for extensive case files. The system can deliver structured outputs and consistently provides explicit citations.

Every response clearly references the laws, rulings and commentaries used. Users can access the original official source via direct link, consult the full text and inspect the specific sections underlying the answer. This transparency transforms AI output into a verifiable research step rather than a black box.

When the system is unsure, it says so instead of making unsupported claims. The structured analysis process ensures that conclusions remain anchored in validated sources.

Clearly defined scope and continuous improvement

DeepLaw does not access real-time data and does not conduct live web searches. All answers are generated exclusively from verified and explicitly ingested sources. The system does not rely on unlisted databases or general web content, and it intentionally excludes tasks outside the legal domain.

An internal benchmark continuously evaluates both the model and each processing stage. In addition, users can submit feedback directly through the platform. Flagged cases are reviewed by legal and AI experts, and validated findings are incorporated into iterative improvements of the processing and evaluation process.

DeepLaw is therefore not simply a general AI adapted for lawyers. It is a secure, jurisdiction-specific legal intelligence system designed to meet the professional, technical and regulatory standards of Swiss legal practice.