/postgres-and-pgvector

Primary LanguagePython

Building AI-Powered Search with PostgreSQL and Large Language Models

Timings

  • 9:30 - 11:00: Session 1
  • 11:00 - 11:15: Coffee Break
  • 11:15 - 12:45: Session 2
  • 12:45 - 1:45: Lunch Break
  • 1:45 - 3:15: Session 3
  • 3:15 - 3:30: Tea Break
  • 3:30 - 4:30: Session 4

Course Overview

This two-day course offers a practical, hands-on journey into building modern AI search applications. We'll explore how to leverage the power of vector embeddings and Large Language Models (LLMs) directly within a PostgreSQL database using the pgvector extension.

Our labs will progressively build your skills, starting with the fundamentals of generating and storing embeddings, and culminating in a final project where you'll build a complete AI-powered search and question-answering system for customer support tickets. By the end, you'll have the confidence to design and implement sophisticated, efficient, and scalable RAG (Retrieval Augmented Generation) pipelines.

Course Modules

This course is divided into modules, each corresponding to a numbered directory in this repository.

  • 02: Getting Setup: An introduction to the core concepts of vector embeddings, pgvector, and the RAG pipeline. We'll also get our local Docker development environment up and running.

  • 03: Generating & Storing Vectors: We'll use a local Ollama model to generate high-quality embeddings and learn how to store them efficiently in our PostgreSQL database.

  • 04: Chunking PDFs: This module covers the critical process of chunking large documents. We'll build a Python pipeline to ingest PDFs, split them into manageable chunks, and store them with page metadata.

  • 05: Querying with Vectors: Learn the fundamentals of vector similarity search. We'll use pgvector operators (like Cosine Similarity and L2 Distance) to find the most semantically relevant documents for a user's query.

  • 06: Querying with LLMs (RAG): We'll build a complete Retrieval Augmented Generation (RAG) pipeline to enhance search results and answer questions using an LLM, based on the context retrieved from our database.

  • 07: JSON and JSONB: Explore how to leverage PostgreSQL's powerful JSONB capabilities to store and query semi-structured metadata alongside vector embeddings for more precise, filtered searches.

  • 08: Hybrid Querying: Go beyond pure vector search by implementing hybrid techniques. We'll combine traditional keyword-based full-text search with semantic vector search for the best of both worlds.

  • 09: Final Lab - AI Ticket Support System: In this capstone project, we'll bring everything together to build a complete AI-powered search application that suggests solutions to new support tickets based on historical data.