![\"gemini-embedding2-retrieval_52](\"https:\/\/storage.googleapis.com\/gweb-developer-goog-blog-assets\/images\/gemini-embedding2-retrieval_52_2.original.png\")
<\/p>\n
Final week, we introduced the Basic Availability (GA) of Gemini Embedding 2 through the Gemini API<\/a> and Gemini Enterprise Agent Platform<\/a>. It\u2019s the primary embedding mannequin within the Gemini API that maps textual content, photographs, video, audio, and paperwork right into a single embedding area, supporting over 100 languages.<\/p>\n On this submit, we are going to discover the various use instances this unified mannequin unlocks, from agentic multimodal RAG to visible search, and present you precisely tips on how to begin constructing them.<\/p>\n The mannequin handles an expansive vary of inputs in a single name: as much as 8,192 textual content tokens, 6 photographs, 120 seconds of video, 180 seconds of audio, and 6 pages of PDFs. By mapping completely different modalities in the identical semantic area, builders can construct various experiences that \u201csee\u201d and \u201chear\u201d proprietary information.<\/p>\n<\/div>\n The true energy of Gemini Embedding 2 is its potential to course of interleaved<\/i> inputs\u2014akin to a mix of textual content and pictures\u2014in a single request:<\/p>\n<\/div>\n \n Python\n <\/p>\n<\/div>\n This allows a extra correct, holistic understanding of advanced, real-world information. In case you want separate embeddings for particular person inputs as a substitute of 1 aggregated vector, use the Batch API<\/a> (help coming quickly for Agent Platform<\/a>).<\/p>\n Multimodal embeddings allow AI brokers to execute multi-step reasoning duties, akin to scanning lots of of recordsdata to repair a codebase or cross-referencing disparate PDFs, with improved understanding and accuracy.<\/p>\n To construct these pipelines with the Gemini API, you should use activity prefixes<\/a> based mostly on the agent\u2019s purpose. These prefixes optimize the ensuing embeddings on your particular activity, serving to the mannequin bridge the hole between brief queries and lengthy paperwork:<\/p>\n<\/div>\n \n Python\n <\/p>\n<\/div>\n Making use of these prefixes at each index time and question time can considerably enhance retrieval accuracy.<\/p>\n Many customers are already seeing a constructive affect from adopting Gemini Embedding 2. Harvey<\/a>, a authorized analysis platform for legislation corporations and enterprises, has seen a 3% improve in Recall@20 precision on legal-specific benchmarks in comparison with their earlier embeddings, resulting in extra correct citations and solutions for legislation corporations and enterprises.<\/p>\nAbout Gemini Embedding 2<\/b><\/h2>\n
from google import genai
\nfrom google.genai import varieties
\n
\nconsumer = genai.Consumer()
\n
\nwith open('canine.png', 'rb') as f:
\n image_bytes = f.learn()
\n
\nconsequence = consumer.fashions.embed_content(
\n mannequin='gemini-embedding-2',
\n contents=[
\n \"An image of a dog\",
\n types.Part.from_bytes(
\n data=image_bytes,
\n mime_type='image\/png',
\n ),
\n ]
\n)
\n
\nprint(consequence.embeddings)<\/code><\/pre>\nAgentic retrieval-augmented era (RAG)<\/b><\/h2>\n
# Generate embedding on your activity's question:
\ndef prepare_query(question):
\n return f\"activity: query answering | question: {content material}\"
\n # return f\"activity: truth checking | question: {content material}\"
\n # return f\"activity: code retrieval | question: {content material}\"
\n # return f\"activity: search consequence | question: {content material}\"
\n
\n# Generate embedding for doc of an uneven retrieval activity:
\ndef prepare_document(content material, title=None):
\n if title is None:
\n title = \"none\"
\n return f\"title: {title} | textual content: {content material}\"<\/code><\/pre>\n