Function Calling for LLMs
summary
Claims that Retrieval Augmented Generation can query tabular warehouses and APIs by using function calling returns JSON to pick and invoke SQL-like functions, then feeding results back in a second LLM call; implements this with Hopworks feature groups and views, LangChain, a Hermes fine‑tuned Llama 3 8B model (reported ~80–90% on a leaderboard) and notes SQL generation benchmarks sit around 65–67%. Demonstrations include an open‑source air quality predictor and a PDF query pipeline, which run end‑to‑end but sometimes pick the wrong function or hallucinate.
AI-written summary · about
chapters (13)
- 0:00RAG with tabular data
- 3:37What function calling LMs do
- 6:47Vector databases vs RAG
- 9:34SQL from natural language
- 11:41Metadata for function calling
- 14:54Pre-defined queries for retrieval
- 17:23Feature views for multiple tables
- 20:35Credit card fraud prediction
- 24:25FTI pipeline for AI systems
- 27:35Live demo: air quality prediction
- 29:59Vector indexes with function calling
- 32:45Recap and Hopsworks
- 33:47Resources and getting started
related talks
15mFrom Words to Data: Using LLMs to Turn Questions into SQL
36mTechniques to avoid common LLM pitfalls
10mKeynote: Unlocking Corporate Data with Retrieval-Augmented Generation
51mAvoid common LLM pitfalls
35mBuilding LLM-based Agents
40mLLMs & Data Warehouses: Reliable AI Agents for Business (Keynote)
24mConversational Search with LLMs
38mLLMs for Self-serve Analytics - Lessons Learned