Scientific foundations
Research-driven engineering, the pivotal paper, the CETIC study
Research-driven engineering: when engineering rests on science
BrainDup invents nothing in a vacuum. Every building block of the framework is a concrete implementation of published, peer-reviewed research. That is not a methodological footnote: it is what separates a production framework from a prototype.
Why does it weigh so much? For three reasons.
- Reproducibility. A technical choice resting on a scientific paper can be checked, reproduced and contested. No black magic in the pipeline.
- Credibility. In front of a decision maker, a technical partner or an auditor, every component of BrainDup points to an identifiable publication.
- Durability. Fashions pass. Scientific results remain. Building on the state of the art means the foundations will hold.
The pivotal paper: GraphRAG Agent with Neo4j and Milvus
In September 2024, Enzo published a decisive technical article on the Neo4j blog: a description of a hybrid architecture combining Neo4j for structured relationships and graph context with Milvus for semantic vector search. The central idea is simple but powerful: the vector alone is not enough. To answer complex questions precisely, you also need to understand the relationships between concepts.
That is BrainDup's technical starting point. The first implementation came with MethodMIND, a semantic search engine over a scientific corpus, built on Milvus alone. The next step was PST2RAG, where the architecture became genuinely hybrid: a dual schema pairing Milvus (vectors) with Neo4j (a knowledge graph). Production maturity came with IA-ACTES.
BrainDup is the industrial implementation of that architecture. The framework extracts, generalises and makes reusable what has been validated in production on real cases.
External validation: the CETIC study
In April 2025, Leandro Collier published a study at CETIC, the Belgian centre of excellence in information and communication technology, titled “Au-delà du simple chatbot”. That independent analysis compares classic vector-based RAG approaches with GraphRAG architectures.
Accuracy
| Context | GraphRAG | Vector-only RAG |
|---|---|---|
| General | 81.67 % | 57.50 % |
| Industry | 90.63 % | 46.88 % |
Efficiency
- 97 % less token consumption. The graph makes it possible to target the relevant information precisely, instead of sending massive blocks of text to the language model.
- Better traceability of the reasoning behind answers. Every answer can be tied back to the graph nodes and relationships that produced it.
Those figures validate exactly the architecture BrainDup implements: vectors plus graph is not a technical luxury, it is a requirement of accuracy.
From science to engineering
Every technical choice in BrainDup rests on at least one published scientific paper.
| BrainDup technical choice | Scientific paper | What it proves |
|---|---|---|
| Modular RAG architecture | Modular RAG: Transforming RAG Systems into LEGO-like Reconfigurable Frameworks (arXiv 2407.21059) | BrainDup implements the LEGO-like frame the research recommends |
| Milvus and Neo4j dual schema | Knowledge Graph RAG for LLM-based Recommendation (arXiv 2501.02226) + étude CETIC | The graph doubles accuracy (81.6 % against 57.5 %) |
| Autonomous AI agents (BMAD) | Agentic RAG: A Survey on Agentic RAG (arXiv 2501.09136) | Agent orchestration is the 2025 state of the art in RAG |
| GPU document conversion (Docling) | Docling Technical Report (arXiv 2408.09869) + Docling Toolkit (arXiv 2501.17887) | Reliable conversion of complex documents |
| Hierarchical RAG over long documents | BookRAG (arXiv 2512.03413) | A hierarchical structure for long documents (a 124-page specification, a legal corpus) |
| Metadata-enriched RAG | Metadata-Driven RAG for Financial QA (arXiv 2510.24402) | Metadata improves relevance |
| Structured prompt engineering | The Prompt Report (arXiv 2406.06608) | Prompting techniques systematised in the BMAD agents |
| Semantic search over a scientific corpus | Searching for Best Practices in RAG (arXiv 2407.01219) | The MethodMIND pipeline follows the identified best practices |
| Large-scale meta-analysis | PERELMAN (arXiv 2512.21727) | An extraction and aggregation pattern applicable to new domains |
This table is BrainDup's scientific traceability map. For every component you can go back to the source, check the results, and understand why that approach was chosen.
What that means in practice
BrainDup is not an experimental prototype built on hunches. It is a framework where every component is scientifically justified, from the dual vector and graph architecture validated by CETIC, through to the autonomous agents described in the most recent surveys.
When you deploy BrainDup, you are not betting on an emerging technology. You are implementing a documented, tested state of the art, already proven in production.
Research-driven engineering is not a slogan. It is a method. And it is what makes BrainDup a base you can build on with confidence.
Take the step
Thirty minutes to assess what sovereign AI can do in your organisation. No commitment.
Free sovereignty review