Enterprise Knowledge Search
Shell needed a domain-specific AI assistant for chemistry and R&D knowledge, but its vast archive of scientific documents required data curation, domain adaptation, and faster training than generic RAG and open-source frameworks could provide.
Reported impact
Shell reported a 30% increase in benchmark accuracy for the custom LLM, a 20% reduction in training time versus other open-source frameworks, and a path to multimodal knowledge support that can further improve decision-making and productivity across R&D workflows.
Shell built a custom chemistry-focused LLM workflow on NVIDIA NeMo to improve knowledge retrieval and document understanding for its R&D teams. The project began by curating 300,000 technical documents collected over decades and filtering them down to 154,000 high-quality documents using exact and fuzzy deduplication, quality filters, language detection, and domain classification. Shell then used the NeMo Framework for domain-adaptive pretraining and supervised fine-tuning so the model could learn Shell's scientific vocabulary rather than rely on generic web language. NeMo's parallelism techniques reduced training time while keeping the workflow on enterprise data. The resulting chatbot is designed to answer questions over Shell's internal chemistry corpus, support document analysis, and evolve through ongoing evaluation and retraining, with future multimodal support planned for images and video.
Technical documents are cleaned and deduplicated with NeMo Curator, selected domain content is fed into domain-adaptive pretraining and supervised fine-tuning in the NeMo Framework, and the resulting model serves chemistry and document-analysis queries for Shell staff.
Initially, Shell had access to 300,000 technical documents collected over decades. These documents were curated down to 154,000 high-quality documents.
Shell accelerated its AI model training time by 20% with NeMo's parallelism techniques.
Shell's custom LLM, trained with NVIDIA NeMo, achieves a 30% increase in accuracy on a custom benchmark.
With the AI-powered chatbot developed by Shell, technology staff would have the possibility to quickly access detailed chemical documents and data.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
NVIDIA Sample Sep 30, 2026 EPRI EPRI needed to run advanced AI for sensitive utility documents and weather-planning workflows locally, because energy-sector research data could not leave controlled environments and shared compute delays limited how… 150 tokens/sec aggregate across 10 concurrent Power Chat users DGX Spark · Dell Pro Max with GB10 · NVIDIA CBottle NVIDIA Sample Aug 3, 2026 StudyFetch StudyFetch needed to keep AI tutoring affordable while scaling live lecture transcription and personalized learning for millions of students, but off-the-shelf cloud transcription created cost and latency ceilings that… 100M+ learning interactions NVIDIA H100 GPUs · NVIDIA L40S GPUs · NVIDIA NIM NVIDIA Sample Jul 7, 2026 Public Ministry of Rio Grande do Sul (MPRS) MPRS needed to manage rising legal caseloads, archived police inquiries, and public-service requests more efficiently, but the volume of legal documents and the need for compliant, locally controlled analysis made… 497 municipalities Archiving Intelligence Assistant (AIA) · Citizen Access Agent · Justice Intelligence platform 