REACH-Compliant AI: An Architecture for a RAG System in the Chemical Industry

The problem: plausible but wrong answers

A practical example: an international chemical company is preparing for an audit. An employee looks up the labelling requirements that apply to a substance in Germany. The AI assistant answers convincingly but mixes US and European requirements.

The result is uncertainty, additional verification work and, in the worst case, violations of applicable law. General-purpose language models produce text that sounds coherent without guaranteeing that it is correct. Where errors affect safety and legal compliance, that is not enough.

The approach: retrieval-augmented generation

Retrieval-augmented generation connects a language model to a verified knowledge base. Instead of relying on training data alone, the system retrieves the relevant documents for each question and formulates the answer on that basis. Every statement can thus be traced back to a source.

For REACH, however, a standard setup is not sufficient, because:

• Regulations differ by country and jurisdiction.
• Substances, mixtures and articles are hierarchically related.
• Technical terms are precisely defined and often used differently in everyday language.

The architecture at a glance

1. Enrich documents with metadata

Each document receives structured metadata such as jurisdiction, country, document type and period of validity. This allows search results to be narrowed down to the relevant legal situation.

2. Interpret questions precisely

An upstream query agent recognises which jurisdiction and point in time a question refers to and translates it into a precise search query with matching filters.

3. Model relationships in a knowledge graph

Relationships between substances, mixtures and regulations are additionally stored in a knowledge graph. This makes it possible to answer traceably which substances a mixture contains and which rules apply to a specific substance.

4. Adapt models to the domain language

Retrieval and language models are adapted to chemical and regulatory texts. This reduces misunderstandings around precisely defined terms.

Together, these components create a system that not only finds information but places it in the right legal and technical context.

Human expertise remains essential

Even the best architecture does not replace professional responsibility. Regulatory affairs experts are an integral part of the system:

• They review answers on a sample basis, and fully for critical questions.
• They maintain a reference dataset of verified questions and answers against which quality is measured.
• Their feedback flows into the further development of retrieval, metadata and models.

Approach in three phases

1. Build the data foundation: capture, clean and add metadata to documents.
2. Build the system: develop and integrate retrieval, query agent and knowledge graph.
3. Test and improve: trial the system in daily work, evaluate it against the reference dataset and optimise it with feedback from the business.

Conclusion

A carefully designed RAG system helps companies in the chemical industry answer regulatory questions faster, more precisely and traceably. What matters is a clean data foundation, an architecture that takes jurisdiction into account and the continuous involvement of domain experts.

We support such projects from architecture to production. Learn more on our pages about RAG systems, Team as a Service and software development.