1. Objective
- Streamline: complex, error-prone manual data entry
- Reallocate: engineering talent to high-value innovation
- Automation: Achieve end-to-end automation for component administration
2. Challenges
- Interpreting ambiguous engineering terminology
- Requiring specialized engineering knowledge
- Handling inconsistent specification formats
- Parsing complex engineering graphics
3. Our Solution
3.1 Solution Overview
- Developed multi-phase data processing to enhance engineering graphic recognition
- Implemented AI-driven engineering data collection and analysis
- Integrated on-premise and public cloud resources
3.2 Intelligent Processing Pipeline
Transforms unstructured engineering content into structured, multimodal data
3.2.1 Data Preprocessing
- Unifies text, tables, and images into a single, context-rich JSON for LLM
- Applies OCR to extract and map image text to its corresponding parent page
- Loads customizable, component-specific attribute lists for direct extraction
3.2.2 Retrieval Augmentation
- Implements Hybrid Search (keyword + vector) for high-precision retrieval
- Utilizes Parent Page Retrieval to preserve critical context
- Employs a cross-encoder reranker to optimize semantic similarity and result relevance
3.2.3 Context Engineering
- Utilizes one universal prompt to handle all component types
- Embeds engineering terminology intent directly into the model
- Pinpoints correct attributes by component model name (especially dimensions)
3.3 Data-Oriented Hybrid Cloud
- On-premise AI platform powered by OCP Grand Teton servers (8x NVIDIA H100 GPUs)
- Software stack built on NVIDIA AI Enterprise, running NVIDIA Dynamo
- Intelligently escalates high-demand workload to public LLM models (GPT, Gemini, Claude)
- Architecture balances data security, cost, and peak performance via intelligent distribution
4. Key Achievements
- Efficiency: 83% reduction in component processing time (2 hours to 20 minutes)
- Autonomy: Fully autonomous, end-to-end AI agent for component database automation
- Scalability: Scalable engineering documents pipeline for knowledge distillation
- Reliability: Hybrid infrastructure for non-stop service
References