Conversational Agents at Scale: Retrieval and Generative approaches
summary
Compares a retrieval-based dialogue management pipeline (intent/entity slots, controlled replies, fast <2s responses, ~150 retrains/day) with a sequence-to-sequence with GRUs generative model (learns from 2–2.5 years of conversations, needs GPUs, handles unseen utterances, even acquires Hindi), and the production choices used: Python/Django REST APIs, models loaded in memory or Redis, TF-Serving for versions, and manual filtering layers. Reports 100+ enterprise bots, deployment at ~30M devices and ~1B interactions; retrieval needs little data, generative needs much.
AI-written summary · about
chapters (16)
- 0:00Conversational Agents at Scale
- 0:26What are Conversational Agents?
- 1:452 approaches to building
- 3:42Retrieval Pipeline
- 6:01Dialogue Management System
- 10:53Word Embeddings
- 16:58Learning
- 19:29Retrieval approach
- 22:56Generative approach
- 24:59Seq2Seg Architecture
- 29:02Seq2Seg approach
- 31:42Hybrid Architecture
- 34:19Things to define
- 36:20ML Model as an API
- 38:23Updating and Swapping model
- 39:32Automate!
related talks
15mRAG Beyond Chatbots
30mConversational AI: Building clever chatbots
14mThe Retail Revolution: Real-World Generative AI Applications
53mGenerative AI: Overview and use cases
33mBuilding the next-generation Conversational AI with Python and Deep Learning
15mHarnessing the Power of Generative AI for Intelligent Applications
48mAn Introduction to Retrieval Augmented Generation
16mTransform customer experience with an array of AI agents