BGE-M3

BGE-M3 is a multilingual, multi-granularity, multi-functional Embedding model open-sourced by the Beijing Academy of Artificial Intelligence, supporting more than 100 languages.

BGE-M3 was developed by the Beijing Academy of Artificial Intelligence (BAAI); the three M's in its name stand for Multi-Linguality, Multi-Functionality, and Multi-Granularity, and it supports more than 100 languages.

Benchmarking on a frozen snapshot of 517 facts with 79 queries (55 exact, 12 paraphrased, 6 third-party relevant, 6 with no expected answer, with hit metrics based on 73 answerable queries): without embeddings and without keywords, overall Hit@1 was 84.9% and paraphrase@1 was 33.3%; without embeddings and with keywords, 90.4% and 66.7%; with embeddings and without keywords, 87.7% and 41.7%; with embeddings and with keywords, 91.8% and 66.7%.

BGE-M3 local embeddings (a 1024-dimensional multilingual model) were added via Ollama, using a graceful degradation mode: a gate observes the lexical coverage of top candidates to decide whether to invoke the model.

Lexical coverage measurements showed: exact match 1.00, paraphrase 0.44, third-party question 0.63; with the threshold set at 0.9, queries that genuinely need semantics open the semantic channel.

On a shared server with a Xeon E5-2450 v2, 15GB RAM, and Debian 13, the measured BGE-M3 cold load took 2.66 seconds against an old timeout of 2.50 seconds; the per-turn limit was adjusted to 2.8 seconds and embedding batching to 60 seconds.

Sizing backfill by fact count is a poor approximation, since model cost is a function of text; batches began to be limited by text volume, capped at about 8000 characters; a full backfill of about 524 facts was estimated at 3.4 minutes, completed in about 19 batches.

Parameter count does not reliably predict indexing speed: EmbeddingGemma-300M indexes more slowly than the larger BGE-M3 (315.0 vs 433.0 chunks per second).

Qwen3-Embedding, EmbeddingGemma 2, and nomic-embed-text v1.5 are all Apache 2.0 licensed, BGE-M3 is MIT licensed, and all can be used directly in commercial products; EmbeddingGemma v1 is governed by the Google Gemma terms; jina-embeddings-v5-text-small uses the CC BY-NC 4.0 license, and commercial use requires obtaining a separate license.

The project is named mini-Rag, open-sourced on GitHub, with the processing pipeline: BGE-M3 → Dense+Sparse Embeddings → Qdrant → Hybrid Retrieval/RRF → BGE Reranker → Relevance Gate → Context Quality Filter → Context Builder → Local LLM.

The BGE-M3 model can simultaneously generate Dense vectors and Sparse vectors; dense vectors excel at capturing semantic similarity, while sparse vectors preserve precise keyword matching capability.

The Beijing Academy of Artificial Intelligence released the BGE-M3 model in January, and the article considers it to be in the top tier globally.

BGE-M3 follows the MIT open-source license, can be deployed locally, supports vectorization of multiple modalities including images, text, and audio, and is completely free.

The three M's in the name BGE-M3 stand for Multi-Linguality, Multi-Functionality, and Multi-Granularity.

BGE-M3 was developed by the Beijing Academy of Artificial Intelligence (BAAI), supports over 100 languages, can simultaneously generate dense vectors, sparse vectors, and ColBERT-style multi-vector representations, and supports a maximum input length of 8192 tokens.

This project uses the BGE-M3 model for text embedding.

BGE-M3 can simultaneously generate dense vectors, sparse vectors, and ColBERT-style multi-vector representations, with a maximum input length of 8,192 tokens.

BGE和E5系列在MTEB基准测试上表现出色，其中BGE-M3支持多语言、多粒度和多功能检索