A hands-on walkthrough of building a conversational analytics agent using MindsDB, an open-source federated query engine. The agent connects MongoDB and HubSpot as data sources, queries both with a single SQL dialect, and answers natural language business questions. The post also covers verbalized sampling, a training-free prompting technique that combats mode collapse in LLMs caused by typicality bias in RLHF alignment. By prompting the model to generate a distribution of responses rather than a single answer, verbalized sampling recovers 66.8% of the base model's diversity and improves output variety by 1.6–2.1x.
Table of contents
[Hands-on] A federated query engine over entire data Verbalized sampling in LLMs P.S. For those wanting to develop “Industry ML” expertise:Sort: