BGE M3
Neural Network
BGE M3 API is an embedding model that encodes sentences, paragraphs, and long documents into 1024-dimensional vectors.
Specifications
- Input
- text
- Output
- vectors (embeddings)
- Context size
- 8 194 tokens
- Vector dimensions
- 1 024
Prompt cost
(per 1M tokens)
How it works BGE M3?
BGE M3 is an embedding model that converts sentences, paragraphs, and long documents into dense 1024-dimensional vectors, designed for multilingual search and large context handling. Vectors are compared by cosine similarity, making the model a foundation for search, ranking, and classification. In BotHub, the BGE M3 model is connected via a unified OpenAI-compatible API—no VPN or foreign cards required, with pay-as-you-go billing via Russian cards and encrypted requests. This enables semantic search over knowledge bases, RAG using Qdrant or pgvector, ticket clustering, card deduplication, and finding similar documents in corporate archives.Frequently asked questions about BGE M3
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
BGE-M3 from BAAI is an embedding model: it turns text into numerical vectors for semantic search, RAG, clustering, deduplication, and document ranking. It is suitable when you need to search by meaning rather than exact word matches, and when building a knowledge base over your data.
The output is not text, but a fixed-dimension vector—an array of numbers in the API JSON response. These vectors are stored in Qdrant, pgvector, or another vector database and compared by cosine similarity. The model context is 8192 tokens per request.
Reasoning mode is not indicated in our data, and it is not required for embeddings: the model encodes the meaning of the text into a vector in a single pass. Logic over the results is built by your pipeline—for example, a generative model to which you pass the found fragments.
Function calling and structured output are not indicated in our data—these are tools for generative models. However, the API response comes in JSON: an array of numbers for each passed fragment, making it easy to parse in code. For function calling, connect a text model via the same API.
Accepting files other than text is not indicated in our data: you pass a string or an array of strings within 8192 tokens as input. If you need images or audio, switch to a multimodal model—the unified OpenAI-compatible BotHub API allows you to do this without rewriting the integration.
Multilingualism is built into the core of this BAAI model: it is designed to work with texts in different languages, and Russian-language search and RAG scenarios are regularly built on it. However, check the final quality on your own data—run a test sample of queries and compare the results.