Google Embeddings 2 (GE2) Dense Retrieval

Updated 4 July 2026

Google Embeddings 2 (GE2) is a hosted, asymmetric bi-encoder model that uses a 2,048-token context window and explicit task conditioning for query and document inputs.
It achieves superior retrieval performance with a BEIR avg.nDCG@10 of 0.638, highlighting its effectiveness on long and heterogeneous passages.
GE2 demonstrates a practical quality–latency trade-off by offering high-quality embeddings through API access with increased latency compared to local models.

Google Embeddings 2 (GE2) is a benchmark label for Google’s Vertex AI-hosted embedding model, identified in the retrieval study as text-embedding-004, and evaluated as a bi-encoder for multilingual dense retrieval and retrieval-augmented generation (RAG)-style passage retrieval (Cirillo et al., 22 May 2026). In that study, GE2 is characterized by a 2,048-token context window and explicit task-type conditioning for RETRIEVAL_QUERY and RETRIEVAL_DOCUMENT, and it ranks first on every evaluated task, reaching BEIR avg.nDCG@10 $= 0.638$ and IT-RAG-Bench nDCG@10 $= 0.282$ , while exhibiting a median end-to-end latency of 231.6 ms (Cirillo et al., 22 May 2026).

1. Definition and model positioning

GE2 is presented as an API-served embedding model rather than a locally run checkpoint. In the benchmark, its defining properties are a 2,048-token context window and explicit task-type conditioning that changes embedding behavior according to whether the input is a query or a document. The paper therefore describes GE2 as an asymmetric bi-encoder in practice, even though it does not use separate encoder weights. This asymmetry is contrasted with the mostly symmetric open-source baselines—BGE-M3, E5-large, Multilingual-E5-large, LaBSE, and paraphrase-multilingual-mpnet—with E5-family models using the instruction prefixes "query:" and "passage:" as a lighter form of query/document asymmetry (Cirillo et al., 22 May 2026).

The same comparison also clarifies the intended operating regime. BGE-M3 is evaluated only in dense retrieval mode to keep the comparison fair, whereas LaBSE and mMPNet are included as models originally designed for sentence similarity, paraphrase, or bitext alignment rather than passage retrieval. This distinction is important because the paper attributes part of GE2’s advantage to the combination of longer context and explicit query/document conditioning, especially when queries and documents have different surface forms and distributions. This suggests that GE2’s retrieval gains are linked not only to embedding quality in the abstract, but also to a task formulation aligned with asymmetric retrieval.

2. Retrieval formulation and scoring

The evaluation pipeline is a standard offline dense-retrieval setup. Documents are chunked according to a chosen strategy and size, each chunk is embedded, the resulting vectors are indexed with FAISS, each query is embedded, top- $k$ chunks are retrieved by cosine similarity, and the retrieved chunks are mapped back to their parent documents for scoring. Because all corpora satisfy $N \le 5{,}000$ , the study uses exhaustive flat search rather than approximate ANN search, so ranking differences are attributed to the models rather than index approximation (Cirillo et al., 22 May 2026).

The retrieval score is cosine similarity,

$\text{sim}(q,d_i)=\frac{\phi(q)^\top\phi(d_i)}{\|\phi(q)\|_2\cdot\|\phi(d_i)\|_2},$

and the reported metrics are Recall@ $k$ , MRR, and nDCG@ $k$ , primarily nDCG@10. The paper gives nDCG as

$\operatorname{nDCG}@k=\frac{\operatorname{DCG}@k}{\operatorname{IDCG}@k},\quad \operatorname{DCG}@k=\sum_{i=1}^{k}\frac{2^{r_i}-1}{\log_2(i+1)},$

with graded relevance labels $r_i \in \{0,1,2\}$ . In this formulation, GE2 is evaluated as a passage retriever rather than as a general semantic-similarity encoder. A plausible implication is that direct score comparisons with sentence-alignment models can be misleading unless the asymmetry of retrieval is made explicit.

3. Benchmarks, corpora, and protocol

The study uses four BEIR subsets chosen to vary lexical mismatch and document length: FiQA-2018 with 648 queries, NFCorpus with 323 queries, SciFact with 300 queries, and TREC-COVID with 50 queries. A specific caveat is that BGE-M3 could not be evaluated on TREC-COVID because of a tokenization incompatibility, so its BEIR average omits that subset. The paper also introduces IT-RAG-Bench, a synthetic Italian benchmark with 3,200 passages and 640 queries. Its passages comprise 1,200 Italian Wikipedia passages, 800 public-administration FAQ entries, and 1,200 articles from the Italian civil code; relevance is annotated at the document level; 200 queries were manually checked by two native Italian speakers; and inter-annotator agreement is reported as Cohen’s $\kappa = 0.81$ (Cirillo et al., 22 May 2026).

Latency and chunking are treated as first-class experimental variables. Latency is measured as the median over 50 independent single-query runs after 5 warm-up runs, including both embedding computation and FAISS retrieval. All local models are run on Apple M-series CPU hardware, whereas GE2 is accessed through Vertex AI, so its latency includes network and service variability. For chunking, the paper evaluates 5 sizes—8, 16, 32, 64, and 128 whitespace tokens—under 3 strategies: fixed non-overlapping chunks, sliding windows with 50% overlap and stride $= 0.282$ 0, and semantic chunking via TextTiling-style boundary detection using sentence embeddings and a cosine threshold $= 0.282$ 1. Semantic chunks are constrained to $= 0.282$ 2 tokens, with fallback to fixed splitting when the range would be violated. Because IT-RAG-Bench has a mean passage length of about 60 whitespace tokens, the authors explicitly caution that some chunking findings are corpus-specific.

4. Empirical performance and quality–latency trade-offs

Across the four BEIR subsets, GE2 attains average nDCG@10 of 0.638, compared with 0.546 for Multilingual-E5-large and 0.538 for E5-large. On IT-RAG-Bench, GE2 reaches nDCG@10 of 0.282, while Multilingual-E5-large reaches 0.279. The paper treats this 0.003 difference as being within expected benchmark noise, and notes that Multilingual-E5-large slightly exceeds GE2 on Recall@10, with 0.489 versus 0.476. GE2 nonetheless retains a small edge on Recall@1 and MRR, with 0.061 versus 0.051 and 0.259 versus 0.243, respectively. The same benchmark reports especially poor dense-retrieval performance for LaBSE, with BEIR average nDCG@10 of 0.188 and Italian nDCG@10 of 0.189, below every dedicated retrieval model in the comparison (Cirillo et al., 22 May 2026).

Model	BEIR avg.nDCG@10	IT-RAG-Bench nDCG@10
GE2	0.638	0.282
mE5-L	0.546	0.279
E5-large	0.538	0.262
BGE-M3	0.437	0.238
mMPNet	0.243	0.243
LaBSE	0.188	0.189

Per-dataset results show where GE2’s margin is largest. On TREC-COVID, it scores 0.799 against 0.702 for Multilingual-E5-large, a gap of 0.097; on FiQA, it scores 0.582 against 0.438, a gap of 0.144. The authors attribute these gains mainly to GE2’s longer context window on long biomedical passages and to asymmetric query/document conditioning on heterogeneous query–document pairs. This suggests that GE2’s advantage is regime-dependent rather than uniform across all multilingual corpora.

Latency sharply changes the practical ranking. GE2’s median end-to-end latency is 231.6 ms, with standard deviation 102.6 and p95 575.5, whereas the local models remain comfortably below 100 ms. The paper characterizes GE2 as roughly 14 times slower than the fastest local models, and attributes much of this variance to network round-trip time and Vertex AI autoscaling cold starts rather than to embedding computation itself.

Model	Median ms	p95
GE2	231.6	575.5
BGE-M3	30.9	32.1
E5-large	30.9	31.4
mE5-L	31.0	31.8
LaBSE	16.6	16.9
mMPNet	16.6	17.0

5. Chunking behavior, deployment guidance, and caveats

The chunking ablation yields a specific pattern: all six models reach at least 95% of peak nDCG@10 by 32-token chunks on the benchmark corpus; there are no meaningful gains at 64 or 128 tokens; and performance collapses at 8 tokens, where nDCG is roughly 0.023–0.046. Semantic chunking helps only at very small chunk sizes, especially 16 tokens. At that setting, the paper reports gains over fixed chunking of +0.090 nDCG for GE2, +0.075 for Multilingual-E5-large, and +0.063 for E5-large. At 32 tokens and above, semantic chunking ceases to provide measurable benefit on this corpus, which the authors describe as a corpus artifact caused by the short average passage length (Cirillo et al., 22 May 2026).

The practical recommendations in the benchmark follow directly from these measurements. GE2 is recommended when retrieval quality is the top priority, when the corpus contains long passages, when query and document language or style are asymmetric, when API latency and cost are acceptable, or when document embeddings can be computed offline so that only query-time latency matters. Multilingual-E5-large is recommended when multilingual retrieval is needed on short corpora under sub-100 ms latency budgets, and the paper treats it as the best default open-source choice. LaBSE is explicitly discouraged as a default passage retriever: the authors argue that its weak results reflect the structural mismatch of using a symmetric sentence-alignment model for an asymmetric passage-retrieval task rather than a defect in its intended use cases.

Several limitations temper the headline ranking. IT-RAG-Bench is synthetic, its queries are templated, and its corpus is short and stylized; only four BEIR subsets are used; latency is measured only on CPU and without batching; and no bootstrap confidence intervals are reported. The authors therefore warn that nDCG gaps below 0.01 should be treated cautiously. In that sense, GE2’s lead on Italian RAG is operationally modest even though its lead on the selected BEIR tasks is substantial.

Within the supplied literature, “GE2” does not name a single cross-domain concept. In the dense-retrieval benchmark, it denotes Google Embeddings 2 as a Vertex AI-hosted text embedding model for multilingual retrieval (Cirillo et al., 22 May 2026). By contrast, Gemini Embedding is a distinct Google text embedding model built by initializing an embedding encoder from Gemini, using bidirectional attention, mean pooling, and a linear projection, and outputting $= 0.282$ 3-dimensional embeddings with Matryoshka representation learning support for 768- and 1,536-dimensional subspaces. That model is framed as a general-purpose embedding system for classification, similarity, clustering, ranking, reranking, and retrieval, and is reported to achieve state-of-the-art results on MMTEB, XOR-Retrieve, and XTREME-UP (Lee et al., 10 Mar 2025).

The ambiguity extends beyond text retrieval. In the sentence-compression literature, one paper using a Google sentence-compression dataset explicitly does not introduce or use anything called “Google Embeddings (GE2)”; instead, it employs contextual embeddings from Flair and BERT, optionally concatenated with GloVe, in a BiLSTM-CRF architecture for deletion-based compression (Nguyen et al., 2020). In geospatial machine learning, AlphaEarth or Earth Embeddings denote 64-band, 10 m spatial resolution geospatial representations used for downstream height inference with U-Net and U-Net++, and are discussed as compact, reusable geospatial interfaces rather than text retrieval models (Hamoudzadeh et al., 19 Feb 2026). In high-energy physics, “GE2/1” refers to a triple-GEM detector system for the CMS Phase-2 Muon System Upgrade, entirely unrelated to machine-learning embeddings (Butalla et al., 2021).

These distinctions are not merely terminological. They indicate that “Google embeddings” can refer to heterogeneous objects: a hosted retrieval encoder, a Gemini-initialized general-purpose embedding model, geospatial Earth representations, or, in a different literature altogether, detector nomenclature. A plausible implication is that any technical discussion of GE2 must specify domain, benchmark regime, and deployment setting. In the retrieval paper, GE2 has a precise meaning: a high-quality but relatively high-latency multilingual dense retriever whose main empirical strengths emerge on long passages and asymmetric query–document distributions.