Multimodal semantic search
Represent product images, descriptions, videos, audio, and documents in one space so users can retrieve media with natural-language or media-based queries.
Independent tool overview
Gemini Embedding 2 at a glance
Gemini Embedding 2 is Google's generally available multimodal embedding model for representing text, images, video, audio, and PDFs in one vector space for cross-modal search, retrieval, clustering, classification, and recommendations.
Visit the official Gemini Embedding 2 site ↗
- Best for
- Cross-modal search, retrieval, RAG, and recommendations
- Model code
- gemini-embedding-2
- Modalities
- Text, image, video, audio, and PDF input
- Input limit
- 8,192 tokens overall
- Output dimensions
- 128–3,072
- Availability
- Generally available via Gemini API and Vertex AI
- Reviewed
- August 30, 2026
Overview
What Gemini Embedding 2 is
Gemini Embedding 2 is Google's first multimodal embedding model in the Gemini API. It converts text, images, video, audio, and PDFs into numerical vectors in one shared space, allowing a text query to retrieve relevant media or different media types to be compared by meaning. It is intended for developers building semantic search, multimodal RAG, document retrieval, recommendations, clustering, and classification—not for generating text or media.
The stable model code is `gemini-embedding-2`, with up to 8,192 input tokens and configurable output dimensions from 128 to 3,072. Google recommends 768, 1,536, or 3,072 dimensions for the best quality. The model became generally available in April 2026 through the Gemini API and Vertex AI, but production adoption still requires careful evaluation: its vector space is incompatible with `gemini-embedding-001`, multimodal inputs have strict file and duration limits, and any migration requires re-embedding the full corpus.
Use cases
Who Gemini Embedding 2 is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Multimodal RAG
Index text and visual document content for retrieval pipelines that need evidence beyond plain text chunks.
Recommendation systems
Compare the semantic meaning of different content types for catalog, media, and discovery experiences.
Classification and clustering
Group or label content across modalities using a common vector representation rather than maintaining a separate model for each media type.
Teams already using Gemini APIs
Add multimodal embeddings through the Google Gen AI SDK, REST API, Batch API, File Search, or Vertex AI while keeping a familiar Google platform stack.
Capabilities
Core Gemini Embedding 2 features
Unified multimodal vector space
Maps text, images, video, audio, and PDF content into embeddings that can be compared across modalities.
Cross-modal retrieval
Supports workflows such as searching images or video with text and finding related content even when the query and result use different media types.
Multilingual text understanding
Google says the model captures semantic intent across more than 100 languages.
Flexible output dimensions
Returns vectors from 128 to 3,072 dimensions so teams can balance retrieval quality against storage and compute cost.
Automatic normalization
Embeddings requested below 3,072 dimensions are automatically normalized, simplifying cosine-similarity, dot-product, and Euclidean-distance workflows.
Task instructions for text
Supports prompt prefixes and document structures that guide text-only embeddings toward retrieval, similarity, classification, or clustering relationships.
Aggregated multimodal embeddings
Can combine several parts, such as text and an image, into one embedding that represents the overall item.
Separate-item embeddings
Can return distinct vectors when each input is wrapped as a separate Content object, while the Batch API handles larger asynchronous jobs.
Batch processing
Offers asynchronous high-throughput embedding at half the standard paid rate for workloads that do not need immediate results.
Google AI Studio and SDK access
Works through the Gemini API, REST, and Google Gen AI SDKs for Python, JavaScript, Go, and other supported environments.
Multimodal File Search
Can be selected as the embedding model for Gemini API File Search stores that need to index and retrieve text and images.
Vertex AI availability
Google also offers the model through Vertex AI for teams that want Google Cloud project, governance, and enterprise deployment controls.
Process
How the Gemini Embedding 2 workflow works
Step 1
Define the retrieval task
Decide whether the application needs cross-modal search, symmetric similarity, query-to-document retrieval, clustering, or recommendations and define an evaluation set first.
Step 2
Choose input representation
Split media into supported text, image, audio, video, or PDF units and decide whether multiple parts should form one aggregate embedding or separate vectors.
Step 3
Select vector dimensions
Benchmark Google's recommended 768, 1,536, and 3,072 dimensions against retrieval quality, vector-storage size, indexing cost, and latency.
Step 4
Embed and store the corpus
Call `gemini-embedding-2` through the standard or Batch API, retain source IDs and permissions, and store vectors in a compatible similarity index.
Step 5
Embed incoming queries
Use the same model and appropriate text task instructions, then retrieve nearest neighbors while enforcing document- and user-level access controls.
Step 6
Evaluate and monitor
Measure recall, precision, ranking quality, latency, and cost on representative multilingual and multimodal queries, then re-test whenever preprocessing or model behavior changes.
Cost
Gemini Embedding 2 pricing and free plan
Gemini API pricing varies by modality. A free tier is available within rate limits, but Google says free-tier inputs may be used to improve its products; paid-tier inputs are not. Paid Batch processing costs 50% of the standard rates. Vertex AI pricing and enterprise terms should be checked separately.
Free tier
Free within limits
Development access for supported modalities, subject to Gemini API free-tier rate limits and data-use terms.
- Text, image, audio, and video input listed as free
- Input may be used to improve Google products
- Batch processing is not available
- Not a substitute for checking production capacity and privacy requirements
Paid standard
$0.20–$12 per 1M input tokens
Synchronous paid Gemini API usage, with the rate determined by modality.
- Text: $0.20 per 1M input tokens
- Images: $0.45 per 1M tokens, approximately $0.00012 per image
- Audio: $6.50 per 1M tokens, approximately $0.00016 per second
- Video: $12 per 1M tokens, approximately $0.00079 per processed frame
- Paid-tier input is not used to improve Google products
Paid batch
$0.10–$6 per 1M input tokens
Asynchronous processing at 50% of standard pricing for jobs that can tolerate delayed completion.
- Text: $0.10 per 1M input tokens
- Images: $0.225 per 1M tokens, approximately $0.00006 per image
- Audio: $3.25 per 1M tokens, approximately $0.00008 per second
- Video: $6 per 1M tokens, approximately $0.000395 per processed frame
- Batch jobs target completion within 24 hours
Pricing checked . Check current pricing at the source ↗
Assessment
Gemini Embedding 2 strengths and limitations
Where it stands out
- One embedding space can support retrieval and comparison across five input formats.
- The stable model is generally available rather than limited to its original public preview.
- Configurable dimensions let teams reduce vector storage and similarity-compute costs.
- The 8,192-token text limit is useful for longer documents and rich content records.
- Standard and discounted Batch APIs support both interactive and large offline indexing workloads.
- More than 100 supported languages make it suitable for multilingual retrieval experiments.
- Gemini API, File Search, SDK, REST, and Vertex AI access provide several deployment paths.
- Modality-specific published pricing makes initial API cost modeling possible.
What to consider
- Gemini Embedding 2 and `gemini-embedding-001` use incompatible vector spaces, so migrations require re-embedding every stored item and query.
- The model produces embeddings, not generated answers, summaries, captions, or edited media.
- The overall input ceiling is 8,192 tokens, and multimodal requests also have tighter file-specific limits.
- A request supports at most six images, one PDF of up to six pages, 180 seconds of audio, or 120 seconds of video under the documented limits.
- Video analysis samples no more than 32 frames and does not process the video's audio track, so fast events or spoken context can be missed.
- Supported image, audio, video, and document formats are restricted; unsupported assets need preprocessing or conversion.
- Combining multiple parts directly produces one aggregated embedding, which can surprise teams expecting one vector per item.
- The old `task_type` parameter is not supported; text-only tasks need carefully formatted instructions in the prompt.
- Lower-dimensional vectors can reduce cost but may reduce retrieval quality, so dimension choice must be benchmarked on the actual corpus.
- The free tier carries different data-use terms from the paid tier and may not be appropriate for private, regulated, or proprietary content.
- Embedding quality benchmarks do not guarantee good rankings for a particular domain, language mix, media type, or access-control model.
- A production retrieval system still needs chunking, metadata, a vector index, reranking, evaluation, monitoring, and authorization controls outside the model.
Compare
Gemini Embedding 2 alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Data Analysis
Pinecone
Use Pinecone when the immediate need is a managed vector database and retrieval layer; it can store embeddings from Gemini or another model but is not itself a direct model replacement.
Explore Pinecone →Business Operations
Cohere Parse 5
Consider Cohere Parse 5 when the main problem is converting complex documents into structured, machine-readable content before retrieval rather than generating a unified cross-modal vector.
Explore Cohere Parse 5 →Consumer
Gemini 3
Consider Gemini 3 when the application needs multimodal reasoning or content generation; it serves a different role and can consume context retrieved with Gemini Embedding 2.
Explore Gemini 3 →Questions
Gemini Embedding 2 FAQs
What is Gemini Embedding 2?
Gemini Embedding 2 is Google's multimodal embedding model for turning text, images, video, audio, and PDFs into vectors in one shared semantic space.
Is Gemini Embedding 2 generally available?
Yes. Google launched it in public preview in March 2026 and lists the stable `gemini-embedding-2` model as generally available from April 22, 2026.
How much does Gemini Embedding 2 cost?
The Gemini API has a free tier and paid modality-specific input rates. Standard paid rates range from $0.20 per million text tokens to $12 per million video tokens, while Batch processing is 50% cheaper. Check the official pricing page for current limits and terms.
What can Gemini Embedding 2 embed?
It accepts text, PNG or JPEG images, MP3 or WAV audio, MP4 or MOV video with supported codecs, and PDFs. It outputs numerical text embeddings rather than media.
What output dimensions does the model support?
It supports dimensions from 128 through 3,072. Google recommends 768, 1,536, or 3,072 for the highest quality, but teams should benchmark storage and retrieval tradeoffs.
Can Gemini Embedding 2 replace gemini-embedding-001 without reindexing?
No. Google says the two models' vector spaces are incompatible, so existing content and queries must be re-embedded and the index rebuilt.
Does it generate one vector or several for multiple inputs?
Multiple parts placed directly in one content input produce one aggregate embedding. Wrap items as separate Content objects—or use the Batch API—when you need a separate vector for each item.
Can it search video audio and frames together?
Not from one video upload. The documented video path samples up to 32 frames and does not process the video's audio track, so audio should be extracted and embedded separately when it matters.
Is free-tier data private?
Google's pricing table says free-tier input can be used to improve its products, while paid-tier input is not. Teams handling sensitive data should review the current terms and use an appropriate paid or enterprise deployment.
Bottom line
Our Gemini Embedding 2 verdict
Gemini Embedding 2 is a strong fit when a search or RAG product genuinely needs one vector space across text and media. Its broad modalities, multilingual coverage, flexible dimensions, GA status, and published API pricing make it practical to evaluate. The major catch is migration and evaluation work: existing Gemini text embeddings cannot be mixed with the new model, multimodal limits affect preprocessing, and retrieval quality must be proven on representative queries before a full corpus is re-indexed.
Visit Gemini Embedding 2 website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →
Phoenix-4
Phoenix-4 - Tavus' real-time human rendering model with emotional intelligence and active listening
TADA by Hume AI
TADA - Hume AI's open-source TTS model that syncs text and audio one-to-one for zero-hallucination speech
.png)
Raven-1
Raven-1 - Tavus's real-time emotional perception model for AI conversations
Critique
Critique - Microsoft's multi-model deep research tool that pits AI models against each other
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.