01
Retrieve documentation for an assistant.
Split your docs into useful passages, index them and retrieve the relevant ones for a question. Return source links with each passage before a separate language model writes the final answer.
EmbeddingGemma 2
Semantic search and RAG.
EmbeddingGemma 2 is Google’s multimodal embedding model. Run it on AI Fabric to turn content into vectors for semantic search and retrieval-augmented generation (RAG), using a consistent configuration for documents and queries.
€5 in free credits for every new customer.
Plan your deploymentAI Fabric access is opening gradually. GPU configuration is confirmed during onboarding.
Search documents, code and media.
Turn your content into a searchable collection, then retrieve evidence for an assistant. Evaluate the returned passages against questions your users actually ask.
02
Search a code knowledge base.
Index functions with meaningful names, comments and repository metadata. Evaluate queries developers really ask, such as where a retry policy is defined, using manually checked results.
03
Organize a mixed content library.
Explore retrieval across supported media while preserving the original asset type and permissions. Test each query-to-asset combination instead of assuming all modalities have equal retrieval quality.
Model capabilities and upstream setup: EmbeddingGemma 2 publisher reference. Reviewed .
Deployment preparation
Generate EmbeddingGemma 2 vectors in Python.
Use the publisher’s query and document prefixes, and record the model revision, dimensions and normalization policy with the index. An embedding endpoint and a vector database are separate parts of the application.
- Install compatible sentence-transformers and Transformers packages and arrange any required checkpoint access.
- Encode a small reviewed corpus and test queries with known relevant results before processing the full collection.
- Batch the indexing workload separately from interactive query encoding, and measure both throughput and retrieval quality.
from sentence_transformers import SentenceTransformer
encoder = SentenceTransformer(
"google/embeddinggemma-2", device="cuda",
)
documents = [
"title: Retries | text: Retry temporary failures with backoff.",
"title: Uploads | text: Validate file types before processing.",
]
query_vector = encoder.encode(
"How should a worker retry a failed request?",
prompt_name="SearchQuery",
)
document_vectors = encoder.encode(
documents, prompt_name="Document",
)
scores = encoder.similarity(query_vector, document_vectors)
print(scores)Text retrieval example using the publisher’s task prefixes. Keep the same model and output configuration for the corpus and queries; this does not create a persistent vector index.
GPU memory & performanceThe 740M model has a smaller weight footprint than the language models listed here. Batch size, sequence length and loaded media encoders still determine peak runtime memory.
In your applicationRebuild or version the index when the embedding configuration changes. Do not mix vectors from incompatible model revisions or dimensions in the same similarity search.
AI Fabric availability & costRTX A4500 (20 GB) and RTX 2000 Ada (16 GB) are the GPU options. Access, runtime settings and billing are confirmed during onboarding. No Blade latency or throughput benchmark is published for this configuration.
Connect embeddings to a RAG pipeline.
Encode the question, retrieve permitted passages and send them to the answering model with their identifiers. Filter by access permissions in your search layer. The assistant should cite the retrieved evidence and report when it does not answer the question.
01 / InputQuestion → query vector
02 / ProcessSearch your document index
03 / ResultPassages → grounded answer
EmbeddingGemma 2
deployment FAQ.
Is EmbeddingGemma 2 a chat model?
No. It returns representations used for search, clustering and related tasks. A separate model generates the answer after your application retrieves evidence.
Can I reduce vector storage?
The publisher documents smaller output dimensions using truncation and re-normalization. Evaluate retrieval quality at the chosen size and apply it consistently to queries and documents.
Does AI Fabric include my vector database?
This guide covers the embedding computation. Choose and operate the index, metadata filters and retrieval layer as part of your application.
How much does running EmbeddingGemma 2 on AI Fabric cost?
Budget document indexing separately from query encoding. Include re-indexing when the model or vector format changes, and account separately for vector storage and search.
The GPU billing unit and idle policy are confirmed during onboarding. Managed Models API rates do not apply to this deployment.
Checkpoint: google/embeddinggemma-2. Access requirements and licence: EmbeddingGemma 2 publisher reference.