Fundamental AI Concepts ======================= **AI Factory** builds on modern AI concepts, architectures, and technologies to enable advanced data-driven and intelligent applications. This page explains key **industry concepts** that form the foundation of AI Factory’s design and capabilities: - AI concepts - Vector and semantic search - Retrieval-augmented generation (RAG) - AI for databases - AI for infrastructure - Hybrid and Sovereign AI architectures To see how these concepts are implemented in AI Factory and Hybrid Manager, start with: - :ref:`Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads) ` - :ref:`AI Factory terminology ` ``---`` Before you start ---------------- You’ll get the most out of this section if you have: - A basic understanding of **AI workflows** - Familiarity with **vector databases** and **semantic search** - Awareness of **AI pipelines** and **LLM architectures** - Helpful: experience with **AI Factory Concepts** and :ref:`Getting Started - AI Factory 101 ` ``---`` Core AI concepts ---------------- Learning systems and model training ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Learning systems enable models to learn from data and improve performance on specific tasks without explicit programming. Common learning types: - **Supervised learning** — Predict based on labeled data (ex: sales forecasting) - **Unsupervised learning** — Find patterns in unlabeled data (ex: customer segmentation) - **Reinforcement learning** — Learn through interaction and reward feedback (ex: game-playing agents) Infrastructure needs: - Scalable storage for training data - Compute for model training (CPUs, GPUs, TPUs) - Optimized serving infrastructure for inference ``---`` Deep learning (DL) & neural networks ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ DL uses deep neural networks with multiple layers to extract and learn complex patterns. Strengths: - Handles unstructured data (images, text, audio, video) - Powers advanced applications such as image recognition, speech recognition, text generation Infrastructure needs: - High-performance compute (GPUs or TPUs) - Optimized data pipelines and model serving infrastructure ``---`` Natural language processing (NLP) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ NLP applies AI to understand and generate human language. Common use cases: - Text classification - Sentiment analysis - Machine translation - Conversational AI - Intelligent document processing - Content summarization Infrastructure needs: - Large model architectures (Transformers) - Low-latency serving infrastructure for real-time interaction - Data pipelines to feed and fine-tune models ``---`` Large language models (LLMs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ LLMs are advanced NLP models based on deep learning architectures such as Transformers. Key characteristics: - Trained on massive text corpora - Generate coherent, context-aware text - Support chatbots, content generation, code assistants, advanced Q&A LLMs are the foundation for modern Gen AI applications. Infrastructure needs: - Enormous compute and storage for training - GPU-optimized serving for real-time inference - Scalable infrastructure for multi-user workloads ``---`` Embeddings and vector representations ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ **Embeddings** are dense vector representations of data, capturing semantic meaning in numerical form. Applications: - Semantic search - Personalization and recommendations - Anomaly detection - RAG (retrieval-augmented generation) How embeddings are used: - Convert text, images, audio, video, or structured data into vectors - Store in vector databases or Postgres with pgvector - Perform similarity search using approximate nearest neighbor (ANN) algorithms ``---`` Vector databases and semantic search ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Databases optimized to store and search embeddings. Key features: - Fast vector similarity search - High-dimensional vector storage - Scale to billions of vectors - Support for hybrid search (vector + keyword / filter) Common use cases: - Semantic search engines - Personalized recommendations - Knowledge assistants with grounding - Fraud detection and anomaly detection - Enhanced Gen AI pipelines (RAG) ``---`` Retrieval-augmented generation (RAG) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ **RAG** combines vector search with LLMs to generate grounded responses. How it works: 1. User query → converted to embedding 2. Vector search retrieves relevant documents / content 3. Retrieved context is injected into LLM prompt 4. LLM generates a response using both its model knowledge and retrieved context Benefits: - More accurate and grounded AI responses - Real-time, dynamic knowledge integration - Support for domain-specific and compliance-aware AI applications RAG is a core architecture pattern supported by AI Factory. ``---`` AI for databases ---------------- Intelligent database management ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Using AI to optimize database operations: - Automated performance tuning - Intelligent query optimization - Proactive resource management - Predictive maintenance - Self-healing capabilities Goal: **autonomous, highly optimized database infrastructure**. ``---`` In-database AI ^^^^^^^^^^^^^^ Perform model training and inference **inside the database**, reducing data movement. Benefits: - Faster insights and response times - Simplified architecture - Lower latency for in-database applications - Real-time prediction capabilities - Streamlined model operations workflows Example: invoking models from SQL in Postgres. ``---`` Vector databases in Postgres ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AI Factory leverages **pgvector** extension in Postgres: - Native vector type and operations - Integration with AI pipelines and embedding models - Full integration with Hybrid Manager and Model Serving - Enable semantic search and RAG pipelines on **your Postgres data** Brings AI capabilities closer to core data. ``---`` AI for infrastructure --------------------- AI-accelerated hardware (GPUs, TPUs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ AI workloads depend heavily on specialized hardware: - **GPUs** — General-purpose acceleration for training and inference - **TPUs** — ASICs optimized for deep learning - **Other accelerators** — FPGAs, inference chips Benefits: - Orders-of-magnitude speedup for model training - High throughput for production inference - Efficient resource utilization in cloud-native AI clusters AI Factory integrates optimized GPU serving via **KServe**. ``---`` Leveraging cloud compute for AI workloads ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Modern AI workloads rely on cloud compute to scale: - GPU / TPU instances on demand - Managed Kubernetes services for training and serving (ex: Hybrid Manager) - Optimized AI stacks (CUDA, PyTorch, TensorFlow, JAX) - Distributed model training with advanced networking - Hybrid and multi-cloud architectures Goals: - Maximize flexibility and performance - Minimize cost - Enable portability across cloud providers and environments AI Factory supports **hybrid cloud** and **Sovereign AI** patterns natively. ``---`` Hybrid and Sovereign AI ----------------------- Modern enterprises increasingly require **Sovereign AI**: - Data remains in your infrastructure and control - AI pipelines leverage your trusted knowledge and content - Model serving runs in your infrastructure (ex: Hybrid Manager + KServe) - Governance, observability, and compliance are built in AI Factory enables: - AI in Postgres - Pipelines and Knowledge Bases with **your data** - Model Serving in your Kubernetes cluster - Full visibility and control via Hybrid Manager dashboards **Sovereign AI** is a first-class design goal for EDB AI Factory. ``---`` Related concepts ---------------- - :ref:`Fundamental AI Concepts ` - :ref:`Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads) ` ``---`` Next steps ---------- To see how these concepts power **EDB AI Factory**, continue with: - :ref:`Gen AI Builder Concepts (Assistants, Knowledge, Tools, Threads) ` - :ref:`AI Factory terminology ` - :ref:`Getting Started - AI Factory 101 ` ``---``