Базовые повторно используемые компоненты развертывания ИИ

Базовые повторно используемые компоненты развертывания ИИ. Восклицательным знаком обозначены чувствительные компоненты.

В статье McKinsey - A data leader’s operating guide to scaling gen AI. September 12, 2024 (https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/a-data-leaders-operating-guide-to-scaling-gen-ai) приведенный ниже текст представлен в графической форме.


Data sources

• API (!)
• File
• Web
• Relational database management system
• Document database

Data repositories

• Raw data
• Object storage
• Curated data
• Graph database
• Vector database
• Online transactional processing
• Columnar storage

Data services

• Generative AI
• Hallucination checker (!)
• Validation (!)
• Generation (!)
• Source attribution
• Prompt library
• LLM chain and agent framework
• Semantic search and retrieval
• Function calling
• Predictive AI
• Input validation
• Model parameter configuration

Data consumption

• API
• User interface (!)
• Chat
• Multimodal input/output

Processing

• LLMs  (!)
• Change monitoring
• API queries
• PII masking
• Chunking and embedding
• Metadata collection
• Multimodality
• Graph neural network
• Reranking
• OCR and text extraction
• Embedding
• Open-source model
• ETL processing

Data and model governance

• Model performance monitoring  (!)
• A/B testing and experimentation
• Routing
• Model registry
• Accuracy evaluation
• "Versioning" and reproducibility
• Model "explainability"
• Reusable pipelines for training and interence
• Request throttling
• Model tuning and training
• Cataloging
• Automated backup and recovery
• Acess requests
• Versioning
• External sharing

Control center gateway

• Financial operations
• Identity and acess management
• Code management
• Secrets management
• Infrastructure operations
• Monitoring and logging
• Container orchestration
• Scheduling
• Sandbox development enviroment
• Shared-development workspaces
• Workflow management