You have probably heard the term “RAG” in conversations about AI. It is one of those acronyms that gets thrown around at conferences without much explanation, leaving most professionals with a vague sense that it is important but no clear understanding of what it actually means or why they should care.
If you work in a regulated industry, RAG is not just important. It is the difference between AI you can trust and AI you cannot.
The Problem RAG Solves
To understand RAG, you need to understand the fundamental limitation of large language models like GPT-4, Claude, and Gemini.
These models are trained on vast amounts of text data. During training, they absorb patterns, facts, and relationships from this data. After training, the model is frozen. It does not learn new information. It does not update itself when laws change. It does not know about the FINMA circular that was published last Tuesday.
This creates three problems for professional use:
Staleness. The model’s knowledge has a cutoff date. Anything that happened after training is invisible to the model. In law, where a single amendment can change the interpretation of an entire statute, working from stale information is not acceptable.
Hallucination. When a model does not have information about a topic, it does not say “I don’t know.” It generates text that sounds plausible but may be entirely fabricated. It will invent case citations, create fictional regulatory provisions, and present them with the same confidence as verified facts.
No source attribution. Even when the model’s answer is correct, it cannot tell you where the information came from. It cannot point to the specific article of law, the specific court decision, or the specific regulatory guidance that supports its statement. For professionals who need to cite sources, this is a dealbreaker.
RAG solves all three problems through a deceptively simple architecture.
How RAG Works
RAG stands for Retrieval-Augmented Generation. The name describes the process:
Step 1: Retrieval. When a user asks a question, the system first searches a curated knowledge base for relevant information. This is not a keyword search. The system converts the question into a mathematical representation (a vector) and finds documents whose meaning is similar. It also runs traditional text searches for exact matches. The results are combined and ranked by relevance.
Step 2: Augmentation. The retrieved documents are provided to the language model as context. The model is essentially told: “Here is the relevant information. Use this to answer the question.”
Step 3: Generation. The model generates its answer based on the retrieved context, not its training data. Because the answer is grounded in specific documents, the model can cite its sources. Because the knowledge base is continuously updated, the answer reflects current information. Because the model is working from real documents rather than parametric memory, hallucination is dramatically reduced.
Think of it as the difference between asking someone to answer from memory versus giving them the relevant files and asking them to answer from the documents. The second approach is more reliable, and you can verify the answer by checking the files.
What Makes RAG “Regulated”
Standard RAG systems are a significant improvement over raw language models. But for regulated industries, standard RAG is not enough. Regulated RAG adds several critical layers.
Verified source data. The knowledge base must contain authoritative, verified information from official sources. For Swiss law, this means federal legislation from Fedlex, cantonal laws from official cantonal databases, court decisions from official court registries, and regulatory publications from FINMA, SECO, and other authorities. The data cannot be scraped from random legal websites. It must come from the source, and the provenance must be documented.
Structured data, not just text. A legal knowledge base is not a pile of documents. Laws have hierarchical structure: books, titles, chapters, articles, paragraphs, subparagraphs. Court decisions have structured metadata: court, date, case number, parties, legal provisions at issue. Regulatory circulars have version histories, effective dates, and amendment records. Regulated RAG preserves this structure, enabling precise retrieval at the article level rather than the document level.
Citation graphs. Legal and regulatory provisions exist in a web of relationships. A law references other laws. Court decisions cite legal provisions and other decisions. Regulations implement directives. Regulated RAG builds and maintains these citation graphs, so when a user retrieves a provision, they also see the decisions that interpret it, the regulations that implement it, and the amendments that have modified it.
Multilingual alignment. In Switzerland, federal laws exist in three official languages (German, French, Italian), with all versions being equally authoritative. A regulated RAG system must handle multilingual queries, retrieve across languages, and present results in the user’s preferred language while maintaining cross-language alignment.
Change detection. Laws and regulations change. A regulated RAG system must detect these changes, update the knowledge base, and optionally alert users who are working with affected provisions. A system that is updated monthly in an environment where laws can change weekly is not meeting the standard.
Audit trail. Every retrieval, every generation, every source used must be logged. When a professional cites information from the system, they must be able to demonstrate the chain of evidence: what was asked, what was retrieved, what sources informed the answer.
The Technical Stack
For those curious about the engineering, here is what sits under the hood of a modern regulated RAG system.
Vector database. Documents are converted into high-dimensional vectors (numerical representations that capture meaning) using embedding models. These vectors are stored in a specialized database (like pgvector, Pinecone, or Weaviate) that supports similarity search. When a query comes in, the system finds the documents whose vector representations are most similar to the query’s vector.
BM25 index. Vector search captures semantic similarity (meaning), but sometimes you need exact matches. BM25 is a proven text search algorithm that excels at finding specific terms, case numbers, article references, and technical terminology. A good RAG system runs both vector and BM25 searches in parallel.
Reciprocal Rank Fusion (RRF). The results from vector search and BM25 are combined using a ranking algorithm that takes the best of both approaches. Documents that rank highly in both searches are prioritized.
Cross-encoder reranking. The initial retrieval casts a wide net. A cross-encoder model then reranks the results by looking at each document-query pair in detail, pushing the most relevant results to the top.
Contextual embedding. A technique pioneered by Anthropic. Before embedding a document chunk, the system uses a language model to generate a brief contextual description: what document this chunk belongs to, where it sits in the document’s structure, and what it is about. This context is prepended to the chunk before embedding, dramatically improving retrieval accuracy.
Multilingual embeddings. Models like BGE-M3 produce embeddings that work across languages. A German-language query can retrieve French-language documents if the meaning matches. This is essential for Swiss legal work.
Together, these components create a retrieval pipeline that is far more sophisticated than a Google search. It understands meaning, respects structure, handles multiple languages, and produces ranked results with full source attribution.
Why This Matters for Your Practice
The practical impact of regulated RAG is not abstract. Consider a few scenarios:
A corporate lawyer reviewing a cross-border transaction. They need to verify provisions across Swiss corporate law, EU merger regulation, and bilateral treaty obligations. A regulated RAG system retrieves the specific provisions from all three domains, shows the citation relationships between them, and highlights recent amendments. What would take half a day of manual research takes fifteen minutes.
A compliance officer assessing a new FINMA circular. They need to identify every internal policy affected by the change. The RAG system maps the circular’s provisions against the institution’s policy framework, produces a gap analysis, and generates a first draft of required updates. The officer reviews, refines, and approves. Total time: two hours instead of two weeks.
A tax advisor evaluating a restructuring. The restructuring touches corporate tax, VAT, stamp duty, and potentially withholding tax obligations across three cantons. The RAG system retrieves relevant provisions from all applicable tax regimes, flags recent rulings that affect the analysis, and presents a structured overview. The advisor focuses on strategy rather than research.
In each case, the RAG system does the retrieval and synthesis work. The professional provides the judgment. The combination is more powerful than either alone.
Choosing the Right RAG System
Not all RAG systems are equal. If you are evaluating AI tools for regulated work, here are the questions to ask:
- Where does the source data come from? If the vendor cannot tell you exactly which authoritative sources feed their knowledge base, that is a problem.
- How often is the data updated? Monthly is not enough for legal work. Weekly is the minimum. Daily or real-time is the standard to aim for.
- Can you verify the sources? Every output should include clickable citations to the original source document. No citation, no trust.
- Where is the infrastructure hosted? For Swiss regulated industries, Swiss hosting is not optional.
- Is there an audit trail? Can you reconstruct the chain from query to retrieval to output?
- Does it handle multilingual queries? In Switzerland, this is a functional requirement, not a nice-to-have.
Enclava, built by Mont Virtua, is designed around these principles. Verified Swiss legal and regulatory data, continuously updated, multilingual, Swiss-hosted, with full source attribution on every output. If you want to see what regulated RAG looks like in practice, visit enclava.ch.