How Samsara Built an Agentic Framework for GenAI Applications
summary
Samsara builds an agentic, RAG‑centric system that ships a knowledge‑base chatbot in days using LlamaIndex and a modular agent API, runs parallel expert agents (retriever, data query, synthesizer) to cut hallucinations, and enforces an early language detector for multilingual queries. A tracing‑centric stack backed by LinkFuse and orchestration via Dagster pushes traces into Ray clusters for automated evals, feedback mining, and Ray Tune experiments; the deployed beta sees happy customers and experimental inference costs as “remarkably reasonable.”
AI-written summary · about
chapters (11)
- 0:00Introduction to agentic frameworks
- 4:22Why Samsara adopted agents
- 6:32Challenges building chatbots
- 9:41Handling ambiguous queries
- 11:45Multilingual chatbot issues
- 13:28Evaluating chatbot performance
- 15:54Iterative architecture design
- 21:36Tracing-centric feedback loop
- 24:04Running evals with Ray
- 25:28Key insights from development
- 27:54Closing and Q&A
related talks
32mBuilding Multi-Agent RAG Systems
21mHarnessing Technological Innovation & Accelerating Uptake
44mAgentic AI at Sage with AgentCore
36mGENAI-Powered chatbot to increase team efficiency
49mHow to build with ASI tech stack
43mArchitecting the Future: Modern Solution Design with Agentic AI
15mServerless & Agentic AI: Better Together
20mEnhancing RAG with Multi-Agent Integration