Gemini Embedding 001

Gemini Embedding 001 — Text Embedding Model

What is Gemini Embedding 001?

Gemini Embedding 001 is Google's production-grade text embedding model, now generally available via the Gemini API and on Segmind. It converts text into high-dimensional numerical vectors that capture deep semantic meaning — enabling machines to understand what text means, not just what it says.

Trained with Matryoshka Representation Learning (MRL), the model outputs 3072-dimensional embeddings by default, but you can truncate them to 768 or 1536 dimensions with minimal quality loss. It supports over 100 languages and accepts inputs up to 2,048 tokens, making it one of the most versatile embedding models available today.

Gemini Embedding 001 holds the #1 ranking on the MTEB Multilingual leaderboard with a mean score of 68.32 — outperforming the second-best model by +5.09 points.

Key Features

• •3072-dimensional embeddings with MRL-based truncation to 768 or 1536 dims
• •100+ language support — one model for global applications
• •8 task types — optimize embeddings for retrieval, classification, clustering, similarity, Q&A, and more
• •2,048-token input limit — handles long documents and paragraphs
• •Synchronous response — no polling needed, results arrive immediately
• •MTEB #1 — top-ranked multilingual embedding model as of 2025

Best Use Cases

Retrieval-Augmented Generation (RAG): Gemini Embedding 001 consistently retrieves the correct answer over 81% of the time in RAG benchmarks — a 3.6% improvement over competing models. Use RETRIEVAL_DOCUMENT to index your knowledge base and RETRIEVAL_QUERY at query time.

Semantic Search: Move beyond keyword matching. Embed user queries and document corpora to surface contextually relevant results even when exact terms differ. Works across science, legal, finance, and code domains.

Clustering and Classification: Group similar documents or classify text by topic, sentiment, or category. The model's high-dimensional, semantically rich embeddings produce tight clusters with minimal noise.

Multilingual Applications: A single embedding model handles 100+ languages without language-specific fine-tuning. Build cross-lingual search or document matching pipelines with one endpoint.

Anomaly Detection: Compare embedding distributions across your data to surface outliers or detect semantic drift over time.

Prompt Tips and Output Quality

• •Always set task_type explicitly — the right task type improves search relevance by 10–30%. Use RETRIEVAL_DOCUMENT for corpus indexing and RETRIEVAL_QUERY for live search terms.
• •For storage-constrained applications, set output_dimensionality to 768 or 1536. MRL training preserves most semantic information even at lower dimensions.
• •Keep inputs under 2,048 tokens. For long documents, chunk text into overlapping passages of ~512 tokens for best retrieval performance.
• •The model is synchronous — expect low-latency responses suitable for real-time applications.

Gemini Embedding 001 — Usage Guide

Gemini Embedding 001 converts text into 3072-dimensional semantic vectors. Use it to power RAG pipelines, semantic search, clustering, and classification across 100+ languages.

Choosing the Right Task Type

The task_type parameter is the single most impactful setting. Always set it explicitly:

Controlling Output Dimensionality

By default, the model returns 3072-dimensional vectors. If storage or latency is a concern, set output_dimensionality:

• •768 — compact, ideal for large-scale vector indexes with millions of documents
• •1536 — balanced quality/size for most production use cases
• •3072 — maximum fidelity; use when precision is critical

MRL training ensures truncated vectors retain the most semantically important information in the leading dimensions.

Recommended Settings (from testing)

These values were validated during testing and produce reliable results:

{
  "input": "Your document or query text here",
  "task_type": "RETRIEVAL_DOCUMENT"
}

For a RAG pipeline, index documents with RETRIEVAL_DOCUMENT and embed queries with RETRIEVAL_QUERY at runtime.

Input Limits

Max input: 2,048 tokens. For long documents, chunk into overlapping passages of ~512 tokens with a 10–15% overlap to preserve context at chunk boundaries.

What is the difference between RETRIEVAL_DOCUMENT and RETRIEVAL_QUERY?

Use RETRIEVAL_DOCUMENT when embedding content you are indexing (e.g. knowledge base articles). Use RETRIEVAL_QUERY when embedding the user's search query at runtime. Using the correct task type for each role materially improves recall.

Can I reduce embedding size to save storage?

Yes. Set output_dimensionality to 768 or 1536. Gemini Embedding 001 is trained via MRL, which preserves the most semantically important information in the leading dimensions — truncation causes only minor quality loss.

How many languages does it support?

Gemini Embedding 001 supports over 100 languages natively. No language-specific configuration is required.

Is this model suitable for real-time search?

Yes. The API responds synchronously with very low latency, making it well-suited for live autocomplete, real-time document retrieval, and interactive chatbots.

What is the max input length?

2,048 tokens. For longer documents, split into overlapping chunks before embedding.

How does it compare to OpenAI text-embedding-3-large?

Gemini Embedding 001 outperforms text-embedding-3-large on MTEB Multilingual benchmarks, ranking #1 overall with a mean score of 68.32 across diverse tasks including retrieval, classification, and clustering.