Gemini Embedding 2

Gemini Embedding 2 — Text Embedding API

What is Gemini Embedding 2?

Gemini Embedding 2 is Google's state-of-the-art text embedding model, built to convert natural language into dense numerical vectors that capture semantic meaning. With the top MTEB score of 68.16 — outperforming OpenAI text-embedding-3-large (64.6) and Cohere embed-v4 (65.2) — it delivers industry-leading retrieval and similarity accuracy across 100+ languages.

Via the Segmind API, you send a text string and receive a float vector ready for indexing in any vector database (Pinecone, Qdrant, Weaviate, pgvector). Eight task-specific modes let you tune the embedding direction to exactly match your use case — from document retrieval to code search to fact verification.

Key Features

• •SOTA MTEB performance: 68.16 overall; 84.0 on code retrieval; 69.9 on multilingual tasks
• •8 task types: RETRIEVAL_DOCUMENT, RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING, QUESTION_ANSWERING, FACT_VERIFICATION, CODE_RETRIEVAL_QUERY
• •Configurable dimensionality: Output vectors from 256 to 768 dimensions — balance quality vs. storage cost
• •100+ language support: Strong cross-lingual retrieval without translation preprocessing
• •Synchronous API: Immediate JSON response with no polling required
• •Segmind infrastructure: Low-latency, scalable endpoint — no GCP project or quota setup needed

Best Use Cases

Retrieval-Augmented Generation (RAG): Use RETRIEVAL_DOCUMENT to embed your knowledge base and RETRIEVAL_QUERY for user questions. The asymmetric task pairing significantly improves precision over single-task embeddings.

Semantic search: Embed a product catalog, documentation site, or support knowledge base. Users get semantically relevant results even when they phrase queries differently from how the content was written.

Classification and clustering: Feed embeddings into a lightweight classifier or k-means cluster to categorize support tickets, content tags, or customer feedback without fine-tuning a full LLM.

Code search: CODE_RETRIEVAL_QUERY produces embeddings tuned for function signatures, docstrings, and code snippets — ideal for developer tools and IDE assistants.

Multilingual pipelines: With 69.9 MTEB multilingual score, a single index handles queries and documents in different languages without separate per-language models.

Prompt Tips and Output Quality

• •Always pair task types correctly: Index documents with RETRIEVAL_DOCUMENT and query with RETRIEVAL_QUERY. Mixing types degrades recall.
• •Chunk long documents: Split text into 200–500 token segments before embedding. Each chunk should represent one coherent idea.
• •Dimensionality tradeoff: 768 (default) gives optimal quality. Drop to 512 or 256 if your vector database costs or query latency are a concern — vectors are truncated, not re-encoded.
• •SEMANTIC_SIMILARITY is not for retrieval: Use it only when comparing two texts directly (e.g., paraphrase detection, duplicate finding). It underperforms RETRIEVAL_* pairs in RAG pipelines.
• •Batch efficiently: Embed multiple documents in sequence; the model is synchronous and fast enough for real-time pipelines at moderate scale.

How to Use Gemini Embedding 2

Gemini Embedding 2 converts text into dense vectors for search, classification, and similarity tasks. Here is how to get the best results.

Choosing the Right Task Type

The task_type parameter is the most impactful setting. Match it to your use case:

Recommended Settings

Based on real testing, these parameters produce reliable, high-quality embeddings:

{
  "input": "Your document or query text here",
  "task_type": "RETRIEVAL_DOCUMENT",
  "output_dimensionality": 768
}

Use output_dimensionality: 512 or 256 if you need to reduce storage costs or query latency.

Indexing a Document Corpus

Embed each chunk with RETRIEVAL_DOCUMENT, then store the resulting vector alongside the chunk text in your vector database. At query time, embed the user's question with RETRIEVAL_QUERY and retrieve the closest vectors by cosine similarity.

Tips for Better Results

• •Chunk documents into 200–500 token segments. Each chunk should cover one topic.
• •Do not mix task types on the same index — all stored vectors must use the same task type.
• •For multilingual corpora, no preprocessing is needed. The model handles cross-lingual retrieval natively across 100+ languages.
• •For code search, pair CODE_RETRIEVAL_QUERY queries against function/method text embedded with RETRIEVAL_DOCUMENT.

What is the difference between RETRIEVAL_QUERY and RETRIEVAL_DOCUMENT?

RETRIEVAL_QUERY embeds a user's question or search query; RETRIEVAL_DOCUMENT embeds the passages or documents in your index. Always use them as a matched pair — this asymmetric approach is how the model is optimised and produces the best recall.

Can I mix task types when comparing vectors?

No. Cosine similarity is only meaningful between vectors produced with the same task type.

How does output_dimensionality work?

The model produces a full-length vector and then truncates it to your specified size. A value of 768 is the recommended default. Smaller values (256, 512) reduce storage and query latency but may slightly lower retrieval accuracy.

Is Gemini Embedding 2 better than OpenAI text-embedding-3-large?

On the MTEB leaderboard, Gemini Embedding 2 scores 68.16 vs. 64.6 for text-embedding-3-large. The quality gap is meaningful for multilingual workloads and code retrieval.

What vector databases work with Gemini Embedding 2 embeddings?

Any database that accepts float arrays — Pinecone, Qdrant, Weaviate, Chroma, pgvector, Redis, Milvus. Set index dimensions to match your output_dimensionality setting (default: 768).

Does the model support batch input?

The Segmind API accepts a single string per request. For batch workloads, send concurrent requests or loop through your corpus sequentially.