Introduction to Neo4j & GenAI

Generative AI & Large Language Models

Generative AI is a class of algorithms and models that can generate new content, such as images, text, or even music. New content is generated based on user prompting, existing patterns, and examples from existing data.

Large Language Models, referred to as LLMs, learn the underlying structure and distribution of the data and can then generate new samples that resemble the original data.

LLMs are trained on vast amounts of text data to understand and generate human-like text. LLMs can answer questions, create content, and assist with various linguistic tasks by leveraging patterns learned from the data.

Instructing an LLM

The response generated by an LLM is a probabilistic continuation of the instructions it receives. The LLM provides the most likely response based on the patterns it has learned from its training data.

In simple terms, if presented with the prompt "Continue this sequence - A B C", an LLM could respond "D E F".

To get an LLM to perform a task, you provide a prompt, a piece of text that should specify your requirements and provide clear instructions on how to respond.

Precision in the task description, potentially combined with examples or context, ensures that the model understands the intent and produces relevant and accurate outputs.

An example prompt may be a simple question.

What is the capital of Japan?

Or, it could be more descriptive. For example:

You are a friendly travel agent
helping a customer to choose
a holiday destination. Your readers
may have English as a second
language, so use simple terms
and avoid colloquialisms.
Avoid Jargon at all costs.

Tell me about the capital of Japan.

The LLM will interpret these instructions and return a response based on the patterns it has learned from its training data.

Potential Problems

While LLMs provide a lot of potential, you should also be cautious.

At their core, LLMs are highly complex predictive text machines. LLM’s don’t know or understand the information they output; they simply predict the next word in a sequence.

The words are based on the patterns and relationships from other text in the training data. The sources for this training data are often the internet, books, and other publicly available text. This data could be of questionable quality and maybe be incorrect. Training happens at a point in time, it may not reflect the current state of the world and would not include any private information.

LLMs are fine-tuned to be as helpful as possible, even if that means occasionally generating misleading or baseless content, a phenomenon known as hallucination.

For example, when asked to "Describe the moon." an LLM may respond with "The moon is made of cheese.". While this is a common saying, it is not true.

While LLMs can represent the essence of words and phrases, they don’t possess a genuine understanding or ethical judgment of the content.

Large Language Models (LLMs) are often considered "black boxes" due to the difficulty deciphering their decision-making processes. The LLM would also be unable to provide the sources for its output or explain its reasoning.

These factors can lead to outputs that might be biased, devoid of context, or lack logical coherence.

Fixing Hallucinations

Providing additional contextual data helps to ground the LLM’s responses and make them more accurate.

A knowledge graph is a mechanism for providing additional data to an LLM. Data within the knowledge graph can guide the LLM to provide more relevant, accurate, and reliable responses.

While the LLM uses its language skills to interpret and respond to the contextual data, it will not disregard the original training data.

You can think of the original training data as the base knowledge and linguistic capabilities, while the contextual information guides in specific situations.

The combination of both approaches enables the LLM to generate more meaningful responses.