Embedding Models

Now that I’m working with my local LLMs using Python, it’s time to learn embedding so that I can take steps to developing local RAG (Retrieval Augmented Generation) application. This is where a model can access specific local large data sets. Not quite a crazy as training model parameters…

Embedding Models

Embedding models are much smaller than generative models. Embedding models analyze a input and convert it into the voodoo language of model matrix vector math in chunks, and store it in a local vector database. Then a generative model can access the database, search for the specific chunks of data relevant to a prompt, and generate responses based on a large, verified, and specific data set.

Nomic has a great reputation for their open weight embedding models, so I’m going to try theirs first. I’m going to grab:

I will also need an embedding model capable of interpreting images and text. These are slightly larger. I’m having a hard time locating one of these that I can run at the moment…

Questions

While looking for one of these, I noticed that there’s a difference between simple vector databases and what I am seeing called “graph” databases. The graph versions are probably much better at finding connections between ideas that are not obvious rather than finding the single answer in the RAG pipeline? I don’t know, food for thought…