FlagEmbedding (BAAI BGE retrieval toolkit)
cuda, cpu
Fields
id
engines/niche-flagembeddingRepo
FlagOpen/FlagEmbedding
Languages
Python
License
MIT
First release
2023
Design goal
retrieval-onlyembedding-and-rerankingresearch-benchmarkingchinese-first-multilingual
Scheduling
none
Backends
cudacpu
Formats
safetensorspytorch
Notable features
THE MODEL FACTORY, NOT A SERVER. FlagEmbedding is primarily a training, fine-tuning and evaluation toolkit plus a model zoo - it produces the BGE embedding and reranker models that other engines then serve. It ships a lightweight API wrapper for convenience, but the artifact that matters is the weights. Recorded here because the model families it defines are the most widely served retrieval models in existenceIt is the only project in this niche that covers BOTH halves of retrieval in one codebase: bi-encoder embedding AND cross-encoder reranking, with matched families. Most serving projects handle one or the otherBGE-M3 is the notable capability: a single model covering multilingual, multi-functionality (dense/sparse/multi-vector) retrieval in one artefact, which collapses a multi-model pipeline into one forward passThe associated papers are real research, not marketing - LLM-Embedder (arXiv 2310.07554, 'Retrieve Anything To Augment Large Language Models'), BGEmbedding (arXiv 2309.07597, C-Pack packaged Chinese embedding resources), and the bge_embedding paper (arXiv 2311.13534). This is a lab that publishes, and the C-MTEB benchmark is part of the ecosystem12212 stars and 927 forks make it the largest retrieval toolkit in this slice by adoption, and the BGE models are served by essentially every general-purpose engine here including [[vllm-pooling-models]] and [[niche-text-embeddings-inference]]The org moved from BAAI to FlagOpen - the historic BAAI/FlagEmbedding path now 404s on the GitHub API, which will break older docs and CI referencesMIT licensed, created 2023-08-02, last push 2026-08-24, 918 open issues, no GitHub releases
Hardware caveats
NO PUBLISHED SERVING THROUGHPUT. The repo publishes quality benchmarks (C-MTEB, MTEB) and training guidance; it does not publish tok/s, TTFT or any serving performance number. Do not mistake retrieval quality leaderboards for serving performance918 open issues on 12212 stars is a very high absolute count for a repo of this age, and reflects a large research surface rather than a serving productTraining/fine-tuning code assumes a researcher's single-node-to-few-node budget, not a serving fleetThe convenience API wrapper is not a production serving path: no batching policy, no scheduler, no continuous batching, no paged KV. It exists for demos and evaluationIt does not serve what it defines - that division of labour is the single most important caveat for a reader deciding whether to adopt it. Adopt the WEIGHTS and serve them elsewhereChinese-first heritage: excellent on C-MTEB and strong multilingual via bge-m3, but English retrieval leaderboards are where the strongest Western alternatives are most competitive
Best for
Producing or fine-tuning the embedding and reranker models your retrieval stack will actually serve - it is a model factory firstBGE-M3 when you want one multilingual dense+sparse+multi-vector model instead of a three-model pipelineReproducing C-MTEB / MTEB retrieval quality work, since the benchmarks and the models live in the same ecosystemA team that wants both embedding and reranking models under one licence (MIT) and one training codebase
Avoid for
Serving. Serve BGE weights with [[niche-text-embeddings-inference]], [[niche-infinity]], [[vllm-pooling-models]] or [[qdrant-fastembed]] - this repo is not the serving layerAnyone needing throughput numbers - there are none published hereSmall-team model development where a 918-open-issue research repo is more surface than you want to track
notes
The model-side anchor of the embedding niche, deliberately recorded to contrast with the serving-side records: FlagEmbedding defines the weights, [[niche-text-embeddings-inference]] and [[niche-infinity]] serve them, [[qdrant-fastembed]] embeds them in-process, and [[niche-colbert]] covers the late-interaction alternative. The BGE-M3 single-model-multiple-retrieval-modes design is the most consequential idea in this slice. NO PUBLISHED SERVING THROUGHPUT - explicit negative. All metadata from the GitHub API on 2026-10-04. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-08-24 (1.3 months before this record), latest release v1.4.2 published 2026-08-24, 72 contributors (distinct committers including anonymous), open_issues_count 918 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 12212 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set.
Notes
The model-side anchor of the embedding niche, deliberately recorded to contrast with the serving-side records: FlagEmbedding defines the weights, [[niche-text-embeddings-inference]] and [[niche-infinity]] serve them, [[qdrant-fastembed]] embeds them in-process, and [[niche-colbert]] covers the late-interaction alternative. The BGE-M3 single-model-multiple-retrieval-modes design is the most consequential idea in this slice. NO PUBLISHED SERVING THROUGHPUT - explicit negative. All metadata from the GitHub API on 2026-10-04. ||| MAINTENANCE (dated fact, GitHub API, 2026-10-04): last commit 2026-08-24 (1.3 months before this record), latest release v1.4.2 published 2026-08-24, 72 contributors (distinct committers including anonymous), open_issues_count 918 (issues PLUS pull requests - see [[ev-github-open-issues-count-includes-pull-requests]]), not archived, 12212 stars. See [[ev-maintenance-status-is-four-dated-facts]] for how to read this fact set.
Cross-references
This record references no other records — it is a leaf. It is cited by 0.