Augmented Retrieval Products with GAI models
summary
Describes building retrieval‑augmented products that run on Azure OpenAI and Azure Cognitive Search, deploying a prototype called project Guru and a commercial pilot PrivacyAid for German data‑privacy content. Uses vector search combined with keyword filtering, finds GPT‑4 gives better comparisons than 3.5, and fine‑tunes a LLaMA‑7B base with LoRA while keeping a RAG layer for citations. Hallucinations drop but complex queries still fail; automated A/B tests and a small tester group drive iteration, product still in extensive pre‑production testing.
AI-written summary · about
chapters (13)
- 0:00Augmented retrieval with GAI
- 3:21Generative AI use cases
- 6:12Limitations of GAI models
- 8:20Fine-tuning in business fails
- 12:16How RAG works
- 14:00Project Guru proof of concept
- 16:28HR Guru user testing
- 19:55Privacy Aid for GDPR
- 23:48Production testing feedback
- 26:16Retrieval limits
- 27:44Fine-tuning + retrieval
- 31:14Tailored vs off-the-shelf
- 34:21Custom GAI products
related talks
48mGetting started with generative AI using Azure OpenAI Service
24mInfuse Generative AI in your apps using
39mDisrupt and Grow with Microsoft Azure AI
51mThe Next Frontier: Azure OpenAI Models and the Arrival of GPT-4
47mRAG at scale: production-ready GenAI apps with Azure AI Search
50mRetrieval Augmented Generation Beyond Toy Demos
01h 30mOptimizing Retrieval for Agentic Systems
48mDeploy large language models responsibly with Azure AI