Model Name	Size	Threshold	Max Tokens	Matryoshka	Multilingual	External
en-2024-04-24	768	0.47	2048	No	No	No
multilingual-2023-08-16	1024	0.7	512	No	Yes	No
multilingual-2024-05-06	1024	0.4	2048	No	Yes	No
Open AI small	1536	0.5	8192	Yes	No	No
Open AI large	3072	0.5	8192	Yes	No	Yes
Google multilingual Gecko	768	0.55	3072	Yes	Yes	Yes
Hugging Face	N/A	N/A	N/A	N/A	N/A	Yes

Embeddings models

Embeddings are like fingerprints for words or data. They help computers understand the similarities and differences between them, making it easier for machines to perform tasks like understanding language or recognizing patterns.

In a Retrieval-Augmented Generation (RAG) system, the quality of the embeddings you choose directly impacts how effectively the system retrieves and understands information. The embedding model you select determines how well your system can find relevant data, interpret user queries, and generate accurate responses. Choosing the right model ensures that your Knowledge Box can deliver precise and contextually relevant information, which is crucial for maintaining high-quality user interactions and decision-making processes.

The choice of embedding model depends on the languages that will be used in both resources and queries. If you plan to use only English, a monolingual English model would be the ideal choice. For multilingual applications, it’s important to consider the specific languages involved. Most multilingual embeddings support high-resource languages—widespread languages like English, Chinese, Spanish, French, and Japanese. However, if your use case involves low-resource languages—those that are less common and have fewer resources, such as Basque, Welsh, or Irish—you’ll need to choose your embedding model more carefully to ensure adequate support.

Agentic RAG's semantic models

These embedding models run 100% on Agentic RAG's infrastructure. This option ensures that all processes remain within Agentic RAG's secure and controlled environment, optimizing efficiency and security.

en-2024-04-24

Our most up to date English model. Suitable for use cases in which both your queries and resources are only in English.

multilingual-2024-05-06

Our most up to date multilingual model providing strong support for both high-resource and many low-resource languages. Optimized for widely spoken languages.

multilingual-2023-08-16

Our best model for low resource and asian languages.

Trusted External Partner Models

These embedding models run on the infrastructure of our trusted partners. This allows you to leverage the expertise and technological capabilities of otherleaders in the field of artificial intelligence.

Google's Gecko

Google's gecko multilingual embeddings. Their use may result in additional costs, contact our sales department if you want to find out more.

OpenAI 3 small

Multilingual embeddings provided by openai, the smallest and fastest they offer. More information here Their use may result in additional costs, contact our sales department if you want to find out more.

OpenAI 3 large

Multilingual embeddings provided by openai, bigger but with more precision than their small ones. More information here Their use may result in additional costs, contact our sales department if you want to find out more.

Hugging Face

We allow you to use any embedding model from Hugging Face to which you have access. To configure the model, you will need to provide the following information:

Hugging Face Model Configuration

• Hugging Face Endpoint URL
  The URL of the Hugging Face embedding model endpoint. This is a required field and must point to the specific model you wish to use.

• Hugging Face API Key
  Your API Key, which grants access to the desired Hugging Face endpoint. This is also a required field.

• Embedding Vector Size
  The size of the embeddings generated by the model.

Advanced Configuration

• Matryoshka Dimensions
  A list of matryoshka dimensions to use for the embeddings, listed in descending order. If you are unsure of what to enter here, it is safe to leave this field empty.

• Similarity Function
  The similarity function determines how the embeddings are compared. You can choose between DOT or COSINE. If you are uncertain, leave this set to DOT.

• Semantic Threshold
  The threshold value for semantic similarity. This setting determines the sensitivity of similarity comparisons. Adjust this value based on your specific use case or leave it at the default setting (e.g., 0.5).

• Passage Prompt
  The prompt to use for passage embedding. Only set this value if your embedding model requires a specific passage prompt.

• Query Prompt
  The prompt to use for query embedding. Only set this value if your embedding model requires a specific query prompt.