The AI engineering industry tends to overcomplicate stuff when it’s not necessary. I pretty often find myself translating jargon into simple explanations. This guide is my running list of the definitions I wish I had when I started learning about using AI.

Remember: you often don’t need to know all the details of how LLMs work. This is a starting point, and if you’re curious about anything specific you can reach out. I’ll be writing more blog posts on related topics.

LLM (Large Language Model)

A model that generates text based on a prompt. (I’m purposely skipping a lot of context on this one!)

Prompt

The content you give to an LLM to get it to generate something.

Prompt Engineering

The skill of writing good prompts for generative models. Some strategies for this are described below.

AI Engineering

Building useful things with AI. Like being a regular software engineer, while also using LLMs as a tool. All engineers will eventually be AI engineers.

Vector Embedding

Converting words into numbers that an LLM can understand. The numbers themselves will have no meaning to a human, but they represent attributes of the words. For example, words like “king” and “queen” or “cat” and “kitten” will have similar vector representations because they are related.

Vector Database

A special database that stores these embeddings.

Similarity search

Finding similar items in a vector database by comparing their number representations. For example, if you search for “puppy”, you might find “dog” and “kitten” because they have similar vector representations.

RAG (Retrieval-Augmented Generation)

A prompting technique. You can break the term down into three parts:

• Retrieval: Looking up relevant information from a knowledge base, often using similarity search in a vector database.
• Augmented: Adding this information to the prompt.
• Generation: Using the LLM to create a response based on both the original prompt and the added information.

This allows the LLM to learn from proprietary information that won’t be in the LLM already.

Chain-of-thought / Reasoning

Asking the LLM to show its work step-by-step. Some papers have shown that this can improve the output quality.

Tool Use

Letting the LLM use other software tools, like looking things up on Google or using a calculator.

Non-determinism

The LLM’s response will be slightly different each time as it’s based on probabilities. This means the system should be designed to handle different outputs from the same prompt.

Transformers / Attention

The architecture behind how modern LLMs are trained. Unless you’re interested in this, you don’t need to know too much more about it.

Eval

Testing how well an AI system performs at specific tasks. Like test-driven development except this also handles non-deterministic outputs.

Zero-shot

Prompting an LLM without giving any examples. It can work if the task is straightforward.

Few-shot

Prompting an LLM by giving it a few examples of the desired behavior. This gives the model more context to give a better response.

One-shot

Prompting an LLM with exactly one example of the desired behavior.

Fine-tuning

Training an LLM on specific data to make it better at specific tasks by showing it lots of relevant examples.

DPO (Direct Preference Optimization)

Training an LLM to give better answers by showing it which responses humans prefer over time. This is an example of fine-tuning.

Structured Output

Getting the LLM to format its responses in a specific way, to fit to a specific schema. This is useful when writing code that calls LLMs, as you want outputs to fit internal types.

Agent

An AI that can make decisions and take actions on its own. In AI engineering, this often means a system that combines various techniques described above.