programmier.con 2025: Semantic AI: A Language Model & an Embedding Model walk into a bar...

Semantic AI als Erweiterung von Generative AI kann der Schlüssel zur Integration von KI in eigene Lösungen sein. In diesem Vortrag zeigt Christian Weyer praxisnahe Architektur-Patterns und Ansätze für die Nutzung von Large und Small Language Models wie GPT oder Llama sowie Embedding-Modellen in modernen Software-Architekturen. Wichtige Konzepte wie Semantic Routing, Light-weight RAG, Structured Output und Observability werden anhand eines End-to-End-Systems mit mehreren Services und Client-Anwendungen demonstriert. Entwickler und Architekten erhalten einen pragmatischen Überblick über die mögliche Umsetzung in eigenen Projekten.

Transcript

1. Semantic AI: A Language Model & an Embedding Model walk into a bar... Christian Weyer | Co-Founder & CTO | Thinktecture AG | christian.weyer@thinktecture.com
2. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Our journey Models for our software Lightweight RAG Semantic Routing Structured Output / Tool Calling 2
3. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... MODELS FOR OUR SOFTWARE 3
4. Language Models understand and generate semantically rich human language, transforming it into text or structured data for both humans and machines. ⚠ Non-deterministic: same input can lead to different outputs. Embedding Models capture semantic meaning by encoding human language into numerical vector representations, facilitating understanding, comparison, and retrieval for both humans and machines. ✅ Deterministic: same input always results in the same embedding. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... 🫱 🫲 Semantic AI Generative AI 4
5. § Language & embedding models part of end-to-end architectures § Embedding models can be run locally § Optimized for CPU § Language models still hard to run locally § High GPU power § High VRAM § High memory bandwidth Semantic AI A Language Model & an Embedding Model Walk Into a Bar... API-based AI model integrations 5
6. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Classical applications & UIs API-based data Document-based data 6
7. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Language-enabled “UIs” – Talk-to-TT 7
8. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... C4 system context diagram § Various tech stacks § Docker-based distributed system 8
9. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... PATTERN STRUCTURED OUTPUT 9
10. § Tools integration (and more) is being standardized with MCP Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Talking to APIs (Function / Tool calling) 10 “When is CW available for a two-days workshop?” System Prompt (+ employee data) + Schema (for structured output) Web API Availability business logic
11. § Frameworks § Pydantic § Instructor § Methodology § Schema with JSON Mode (not Function Calling) § SLM/LLM § Llama 3.3 70B on Cerebras (very fast) Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Technical implementation – Structured Output 11
12. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... PATTERN LIGHTWEIGHT RAG 12
13. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Talking to documents (Retrieval-augmented generation) Cleanup & Split Text Embedding Question Text Embedding Save Query Relevant Results Question Answ er w / sources LLM Embedding Model Embedding Model 💡 Indexing / Embedding Question Answering .md, .docx, .pdf etc. “Lorem ipsum…?” 💡 Vector DB 13
14. § Frameworks § LangChain § FastEmbed § Lightweight & efficient for generating text embeddings § Embedding model § jinaai/jina-embeddings-v2-base-de (local) – 768 dims § Vector store § PostgreSql (pgvector) vector store § LLM/SLM § Llama 3.3 70B on Cerebras (very fast) Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Technical implementation – Lightweight RAG 14
15. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... PATTERN SEMANTIC ROUTING 15
16. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Semantics-based decisions for user interactions Guarding (e.g. prompt injection) Routing (selecting correct target) “Lorem ipsum…?” Target RAG Target API Call Target … something else … Fine-tuned NLP Model Embedding Model 16
17. Guarding § Frameworks § llm-guard § HuggingFace Transformers § NLP model § deepset/ deberta-v3-base-injection (local) Routing § Frameworks § semantic-routing § FastEmbed § Embedding model § intfloat/ multilingual-e5-large (local) – 1024 dims § Vector store § PostgreSql (pgvector) Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Technical implementation – Semantic Guarding & Routing 17
18. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... Typical Semantic AI patterns & solutions – in end-to-end software engineering Lightweight RAG Structured Output Semantic Guarding & Routing Insightful Observability 18
19. Semantic AI A Language Model & an Embedding Model Walk Into a Bar... 19 AI-based solutions are ≅10% AI and 100% software engineering. 19
20. Thank you! Christian Weyer christian.weyer@thinktecture.com https://thinktecture.com/christian-weyer 20