Prompt Engineering for Generative AI

Chapter 5. Vector Databases with FAISS and Pinecone

This chapter introduces the concept of embeddings and vector databases, discussing how they can be used to provide relevant context in prompts.

A vector database is a tool most commonly used for storing text data in a way that enables querying based on similarity or semantic meaning. This technology is used to decrease hallucinations (where the AI model makes something up) by referencing data the model isn’t trained on, significantly improving the accuracy and quality of the LLM’s response. Use cases for vector databases also include reading documents, recommending similar products, or remembering past conversations.

Vectors are lists of numbers representing text (or images), which you might think of as coordinates for a location. The vector for the word mouse using OpenAI’s text-embedding-ada-002 model is a list of 1,536 numbers, each representing the value for a feature the embedding model learned in training:

[-0.011904156766831875,
 -0.0323905423283577,
 0.001950666424818337,
...]

When these models are trained, texts that appear together in the training data will be pushed closer together in values, and texts that are unrelated will be pushed further away. Imagine we trained a simple model with only two parameters, Cartoon and Hygiene, that must describe the entire world, but only in terms of these two variables. Starting from the word mouse, increasing the value for the parameter Cartoon we would travel toward the most famous ...