EIEB: Enterprise Intelligence Embedding Benchmark

Paper: Does Your Embedding Model Work at Work? Benchmarking Chinese Text Embeddings on Enterprise Documents

What is EIEB?

EIEB is a Chinese benchmark for text embeddings on enterprise documents. It contains 74 datasets that cover 22 document domains in 7 office scenarios and 6 embedding tasks. Within a domain, the same documents serve several tasks, so a difference between tasks is not confounded with a difference between datasets. Building EIEB took about six months, and several departments of an industrial partner already use it to select embedding models.

Why EIEB?

Intranet search, question answering over documents, ticket routing, and candidate matching all rely on text embeddings to retrieve, classify, and match documents. Embedding models are usually compared on public leaderboards such as MTEB and C-MTEB, whose retrieval data comes largely from Web search. In C-MTEB, six of the eight retrieval datasets take their queries from the search logs of Sogou, Baidu, Bing, Taobao, Youku, and Quark. Enterprise documents differ from this data in two ways. They come from specialized domains with their own terminology and relevance criteria, and they range from short tickets to reports tens of thousands of characters long. EIEB asks whether a model that ranks high on Web-derived benchmarks also ranks high on enterprise documents.

Benchmark Overview

EIEB datasets by office scenario (inner ring), document domain (middle ring), and task (outer ring).

Task	Datasets	Records	Metric
Classification	14	193,970 documents	Accuracy
Clustering	14	180,549 documents	V-measure
Retrieval	11	135,444 queries	NDCG@10
Reranking	11	135,444 queries	MAP
STS	12	139,129 triples	Spearman
Pair Classification	12	134,654 triples	Max AP

Documents average 25 characters for office intents and about 49,000 for financial reports. The 11 retrieval corpora have a median document length of 4,947 characters, while the eight C-MTEB retrieval datasets average 31 to 874 characters per document. Each retrieval query has one relevant document in a corpus of 7,601 to 27,456 documents, and reranking adds 100 BM25 hard negatives to each query. STS and pair classification use triples of an anchor, a hard positive, and a hard negative.

Most documents are public Chinese documents: contracts, court judgments from China Judgements Online, patents, bidding announcements, financial disclosures from CNINFO and China Money, and job postings. Resumes, emails, meeting records, and IT tickets are rarely public in Chinese, so we translate public English datasets such as Kaggle and LiveCareer resumes and MeetingBank transcripts. Office intents and procurement tickets are synthesized from examples provided by the industrial partner, and only the synthesized records are released. Qwen3.8-Max is used for translation, label induction, query generation, and triple construction. Every dataset passes a human audit with at least 95% of sampled items judged valid, and every translated dataset reaches at least 99%.

Main Results

We evaluate 39 embedding models from 0.1B to 14B parameters with frozen embeddings. Each task score is the mean primary metric over its datasets, multiplied by 100, and Avg. is the unweighted mean of the six task scores.

• No single model leads every task. Octen-Embedding-8B has the highest average (45.00) and leads STS and pair classification. Nemotron-3-Embed-8B-BF16 leads retrieval and reranking, 7.89 NDCG@10 points ahead of the next retrieval model. gte-Qwen2-7B-instruct leads classification, and F2LLM-v2-14B leads clustering.
• Fine-grained matching is hard. 24 of the 39 models have a negative STS score, and 31 score below 50 in pair classification, the level of random scoring. Most models rank the hard negative above the positive.
• Size helps on average, but small models can compete. The 180M voyage-4-nano outperforms 9 of the 14 models with 7B or more parameters in retrieval.
• Longer inputs help retrieval most. From 512 to 2,048 tokens, retrieval and reranking gain 4.9 and 5.0 points on average, while classification, clustering, and pair classification gain only 0.6 to 1.7.

Best scores are in bold and second-best scores are underlined. Context is the maximum input length stated for each model, in tokens.

Model	Size	Context	Cls.	Clu.	Rerank.	Retr.	STS	Pair Cls.	Avg.
F2LLM-v2-14B	14B	40K	56.40	43.49	41.68	36.49	12.72	52.96	40.62
KaLM-Embedding-Gemma3-12B-2511	12B	32K	56.60	40.94	32.76	24.35	-0.30	48.46	33.80
bge-multilingual-gemma2	9B	8K	54.54	40.28	29.97	23.10	-3.11	47.47	32.04
Qwen3-Embedding-8B	8B	40K	55.85	42.21	47.61	43.10	1.75	49.52	40.01
QZhou-Embedding-Zh	8B	40K	50.58	40.99	13.05	5.77	-2.51	47.87	25.96
Octen-Embedding-8B	8B	40K	54.33	41.38	52.44	48.37	18.29	55.19	45.00
octen-law-8b-v1	8B	40K	54.33	41.30	52.46	48.35	18.24	55.17	44.98
Nemotron-3-Embed-8B-BF16	8B	32K	53.63	37.88	62.05	56.26	2.50	49.30	43.60
F2LLM-v2-8B	8B	40K	56.39	43.36	39.36	34.46	7.08	50.92	38.60
e5-mistral-7b-instruct	7B	32K	52.82	39.77	24.36	16.89	-3.29	48.34	29.82
NV-Embed-v2	7B	32K	38.24	29.35	20.26	13.81	-1.96	48.36	24.68
inf-retriever-v1	7B	8K	55.69	42.06	48.41	43.22	-6.88	46.03	38.09
gte-Qwen2-7B-instruct	7B	8K	57.38	42.49	31.29	23.99	-10.22	46.31	31.87
gte-Qwen1.5-7B-instruct	7B	32K	55.02	41.35	26.93	19.35	-11.14	46.00	29.59
Qwen3-Embedding-4B	4B	40K	54.84	40.93	44.59	39.41	-0.42	48.16	37.92
Octen-Embedding-4B	4B	40K	54.31	41.62	46.54	41.78	12.74	52.81	41.63
Youtu-Embedding	2B	1K	49.29	38.04	26.66	18.47	-7.48	45.90	28.48
BidirLM-1.7B-Embedding	1.7B	32K	54.27	42.85	25.51	19.32	0.55	49.60	32.02
inf-retriever-v1-1.5b	1.5B	8K	52.96	40.86	38.85	32.66	-7.46	45.74	33.94
gte-Qwen2-1.5B-instruct	1.5B	32K	53.53	39.74	31.60	23.57	-2.21	47.76	32.33
BidirLM-1B-Embedding	1B	32K	52.35	40.48	21.50	14.30	0.55	49.29	29.75
Jasper-Token-Compression-600M	600M	32K	52.92	41.55	29.10	21.99	-0.05	48.31	32.30
Qwen3-Embedding-0.6B	596M	32K	51.63	41.19	34.75	27.76	0.47	48.26	34.01
jina-v5-text-small	596M	32K	40.42	40.65	41.20	35.36	7.24	50.93	35.97
jina-embeddings-v3	572M	8K	46.65	42.02	26.09	18.73	5.38	50.90	31.63
bge-m3 (Dense)	560M	8K	47.96	38.19	23.13	13.80	-4.09	47.10	27.68
multilingual-e5-large-instruct	560M	512	50.49	41.47	22.90	15.05	-3.74	47.88	29.01
KaLM-Embedding-V2.5	494M	32K	54.48	42.57	30.58	23.63	0.20	48.87	33.39
KaLM-embedding-multilingual-mini-v1	494M	512	51.39	39.09	25.93	18.18	-6.74	46.62	29.08
Yuan-embedding-2.0-zh	326M	512	49.33	40.43	16.45	8.50	-9.09	45.88	25.25
ritrieve_zh_v1	326M	512	48.59	40.42	16.64	8.25	-5.96	46.85	25.80
Conan-embedding-v1	326M	512	48.40	40.80	16.21	7.81	-7.22	46.54	25.42
xiaobu-embedding-v2	326M	512	48.28	40.28	15.29	7.21	-7.79	46.35	24.94
zpoint_large_embedding_zh	326M	512	48.00	40.40	16.92	8.75	-4.84	47.29	26.09
EmbeddingGemma-300M	307M	2K	49.07	40.17	13.27	3.91	-22.43	42.29	21.05
bce-embedding-base_v1	279M	512	46.31	35.39	16.28	7.42	-16.34	43.97	22.17
jina-v5-text-nano	239M	8K	31.63	37.80	34.93	28.04	0.54	49.21	30.36
voyage-4-nano	180M	32K	54.72	41.10	43.58	37.04	10.04	51.99	39.75
gte-base-zh	102M	512	46.44	37.46	15.24	7.13	-6.48	46.66	24.41

Rankings from C-MTEB Do Not Transfer

(a) Spearman correlations between C-MTEB and EIEB task rankings, with matching tasks on the diagonal. (b) EIEB score lost by choosing the C-MTEB leader.

• Rankings agree only weakly. C-MTEB and EIEB rankings correlate at 0.13 in classification and 0.45 in clustering. In retrieval, reranking, STS, and pair classification, the correlation is near zero or negative.
• The disagreement has a direct cost. Choosing the C-MTEB leader loses 34.7 NDCG@10 points in retrieval and 31.2 MAP points in reranking, relative to the best EIEB model among those with C-MTEB scores.
• No other C-MTEB task is a better proxy. C-MTEB reranking correlates negatively with every EIEB task, and the largest positive correlation in the matrix is 0.45.
• The pattern holds in every scenario. Retrieval and reranking rankings correlate negatively with C-MTEB in all seven scenarios. The scenarios themselves agree at 0.45 to 0.94, so model rankings are more consistent across enterprise scenarios than between any scenario and C-MTEB.

Rankings on Web-derived benchmarks therefore do not reliably carry over to workplace collections. Even within EIEB, the best model changes with the task weights and the document domain, so model selection should look at task-level and domain-level results.

License

EIEB is released under the MIT License. This license applies to the rights held by the EIEB contributors. Third-party source materials remain subject to their original licenses and applicable terms.