Shell Deploys RET-RAG® Semantic Search System for TechXplorer Digest Archive

Title

Shell Deploys RET-RAG® Semantic Search System for TechXplorer Digest Archive

Content

Shell deployed a RET-RAG (Requirement Extraction Tools-Retrieval Augmented Generation) system called RET-RAG® for its TechXplorer Digest magazine archive, developed in collaboration with Huracán Technical Advisory and Energy Transition Campus Amsterdam (ETCA). The system enables semantic search across more than 50 volumes of Shell's internal and external TechXplorer Digest editions, which document Shell's scientific research and technological developments. The RET-RAG® pipeline follows four stages: document ingestion, embedding and storage, query handling, and LLM interface. Documents are uploaded as PDFs or Word files, then automatically parsed by a RET backend to extract text and metadata including headings, page numbers, figure captions, and other contextual data. Text is chunked into segments of approximately 256 tokens each for granular processing. Each text chunk is converted into a 384-dimensional vector embedding using the GTE-large sentence transformer model, then stored in ChromaDB vector database alongside metadata linking back to source documents. When a user poses a query, it is embedded with the same GTE-large model and compared against stored chunks via cosine similarity search, retrieving the top 5 most relevant snippets. The retrieved context chunks are combined with the user's query into an augmented prompt and fed to a local LLM (Gemma 3B, Gemma 27B, or Qwen-7B served via Ollama runtime). The LLM generates grounded responses citing specific TechXplorer Digest issues and page numbers, ensuring authenticity. The entire system runs on Huracán's locally hosted infrastructure with Docker containers, ensuring Shell's data remain private and secure. The RAG approach ensures answers are grounded in actual archive content rather than hallucinated, with post-processing adding inline citations and source links.