Embedding Models

For the FinMAS system to be successful, the relevant data needs to be sent to the LLM agents together with the task that the LLM agents are set to perform. For the system to make a best effort to find the most relevant data for the query, it uses embedding models to convert textual data into dense numerical representations that we call embeddings. The main concept is that by storing the data as numerical vectors, the model would be able to estimate which parts of the data are similar to each other and which part of the data are very different from each other.

The choice of embedding model can significantly affect the result from the analysis done by the Multi-Agent system, as the model is responsible for finding the relevant data to sent to the LLM agent.

OpenAI embedding model

If an OpenAI model is used, then an embedding model of OpenAI will be used. The default embedding model is set to text-embedding-3-small which is most cost-effective option from OpenAI. There is also an option to use text-embedding-3-large, and these embedding models were released early in 2024. We generally do not recommend to use the text-embedding-ada-002 model as it is not as cost-effective.

HuggingFace embedding models

The user can choose from a pre-defined selection of embedding models that are retrieved from HuggingFace. When an embedding model is retrieved from HuggingFace it will be downloaded locally to the directory set in the embedding_models_dir.

The default option is to use BAAI/bge-small-en-v1.5 This model is developed by Beijing Academy of Artificial Intelligence (BAAI) and is a small English text embedding model. It has a maximum sequence length of 512 tokens and outputs a vector representation with a dimension of 384.