Boosting Sales with Intelligent BI Agents
summary
A BI system is presented that turns natural language into validated SQL using Llama 3.1 fine-tuned models, Retrieval-Augmented Generation and vector embeddings, with a graph-backed schema linking layer (neo4j → papygraph) to resolve joins and missing foreign keys. Multi-agent orchestration splits complex requests, runs a multi-stage validator to prevent injection and trim redundant joins, and an optimizer reduces query size and latency. A synthetic training set of 200+ synthetic complex queries plus paraphrases measurably raises SQL accuracy; a live demo returns top-10 customers and a bar chart.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 1:10NL to SQL challenges
- 5:47System architecture overview
- 13:26Fine-tuning LLMs
- 23:26BI agent demo
- 29:27Challenges and future
related talks
15marchive – Manohar Sai Jasti
45mFrom hand-down maps to AI/BI: a journey toward Databricks
15mFrom Words to Data: Using LLMs to Turn Questions into SQL
30mRyoma: Building Open Source AI Agents for Enterprise Data Analysis
38mLLMs for Self-serve Analytics - Lessons Learned
45mWhat’s Next in AI for Data and Data Management
15mGraphRAG-Powered AI Agents: Transforming Corporate Metadata into Intelligent Knowledge
36mBuilding Agentic AI Assistants: Lessons from productionising agents in Preclinical R&D