Using LLMs from Clojure
| This is part of the Scicloj Clojure Data Tutorials. |
1 LLMs and Clojure
LLMs (Large Language Models) are a class of predictive models which can create “content” in various forms, original “text” content.
They are ultimately based on the completion of “text” a user is giving them. The quality of this got so good lately, that we interpret this “text completions” as artificial intelligence, as they imitate with a very high quality what a human might generate.
These models come “pre-trained”, so they have learned a probability distribution of word sequences, which enables them to predict the next word base on any sequence of words.
The inner working of LLMs have as well the concept of embeddings, which means to represent text as high dimensional vectors, where mathematical vector distance is correlated with semantic similarity.
In their most popular form they are presented to users as “chat bots” with which a user can have a a coherent conversion with questions and answers.
This being a “conversation” is an illusion from the technical level. The model itself is stateless, it uses previous parts of the conversation as input for its prediction, which creates the illusion of coherence.
The following chapters show three examples for using LLMs from Clojure:
- a simple chat completion
- using a vector store and embeddings to perform a semantic search
- show case a simple RAG (Retrieval-Augmented Generation) use case