Run EmbeddingGemma 2 locally for text search

A local embedding model at google/embeddinggemma-2 that can score a query against a document. The card's quick start is text. Images and audio are a later step, and they cost extra memory. Start with text.

The launch blog is October 6, 2026. The commands below are copied from the model card, not from a secondary tutorial.

pip install -U sentence-transformers transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("google/embeddinggemma-2")

query = "What causes the northern lights?"
document = "The northern lights are caused by charged particles from the sun."

query_emb = model.encode(query, prompt_name="SearchQuery")
doc_emb = model.encode(document, prompt_name="Document")
print(model.similarity(query_emb, doc_emb))

SearchQuery applies task: search result | query: {query}. prompt_name="Document" applies title: none. If the document has a real title, do not use that prompt name. Format it yourself:

doc_emb = model.encode(f"title: {title} | text: {document}")

Prefixes are for text. The card says to pass images, video, and audio with no prefix.

• Query and document embeddings must use the same dimension. The native size is 768.
• A search query prefix on both sides is the wrong setup. Retrieval is asymmetric: query prefix on the query, document format on the corpus.
• Classification, clustering, and similarity are symmetric. The card's prompt names for those are Classification, Clustering, and SentenceSimilarity. Do not reuse SearchQuery for them.

• Model card quick start: https://huggingface.co/google/embeddinggemma-2
• Launch blog: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/