Improve LLM-based Applications with Fallback Mechanisms
summary
Shows how to harden RAG apps by wiring fallback branches in Haystack: build pipelines of embedder → retriever → prompt builder (Jinja-style templates) → generator, using Hugging Face inference (Mistral-7B-Instruct-v3) or other OpenAI-compatible providers. Implements a conditional router that detects a “no answer” phrase and routes to a web search component or alternative stores (Slack, Notion, Drive), an early-routing pattern where the LLM selects a branch as a tool, and a looped output validator using Pydantic schema with a max_loops limit to autocorrect invalid JSON (notes hallucination risks on missing fields).
AI-written summary · about
chapters (8)
- 0:00Introduction and agenda
- 1:41Understanding RAG pipelines
- 4:37Building with Haystack
- 10:20Conditional routing fallback
- 15:01Early routing strategies
- 21:20Structured output validation
- 28:20Conclusion and resources
- 30:13Q&A session
related talks
35mBuilding LLM-based Agents
01h 00mFrom LLM Theory to Practical Agentic Implementations
01h 39mMastering LLMs: From Prompt Engineering to Agentic AI
50mFrom Prompt Hacking to Architectural Determinism: Engineering Reliable GenAI Systems
40mGo Beyond Basic RAG with Agentic Behavior
31mBeyond LLMs and Maturing RAG use cases
48mBeyond the Prompt: Evaluating, Testing, and Securing LLM Applications
36mTechniques to avoid common LLM pitfalls