For years, building a multimodal semantic search pipeline meant maintaining separate models per data type — one for text, one for image, one for audio — then trying to align those vector spaces with fragile fusion layers. On March 10, 2026, Google ended that era with the launch of Gemini Embedding 2: the first natively multimodal embedding model in the Gemini family, in Public Preview via Gemini API and Vertex AI.

The numbers behind the migration

The problem it solves

Enterprise data is multimodal by nature. Customer support involves text tickets, call recordings, error screenshots and PDF manuals. A product analytics system handles demo videos, written specs and catalog images. Indexing all of this semantically required separate pipelines — and quality dropped at the seam between modalities.

Gemini Embedding 2 fixes this at the foundational layer: by training a single model across all modalities, the vector distance between “an audio recording of a customer complaining about latency” and “a knowledge-base article on performance tuning” is semantically coherent — with no intermediate translation layer.

"The bridge between different media types has finally been built. Use this with complex document similarity tasks, and the results in semantic proximity should be a massive leap forward for RAG pipelines."

What’s in the spec

The gemini-embedding-2-preview model is built on the Gemini architecture and inherits its multimodal understanding. Per-modality limits:

A key architectural detail: the model accepts interleaved input — combine multiple modalities in a single request (text + image + audio) and receive one aggregated embedding that captures the relationship between them. Different from generating separate embeddings and averaging.

Matryoshka: flexibility without sacrifice

Gemini Embedding 2 incorporates Matryoshka Representation Learning (MRL) — semantic information nested hierarchically inside the vector. The first 768 values already hold a useful representation; the next 1536 add nuance; the full 3072 deliver maximum fidelity.

Pick dimensionality at inference time via the output_dimensionality parameter:

from google import genai
from google.genai import types
client = genai.Client()
Reduced-dimension embedding (75% storage savings)
result = client.models.embed_content(
model=“gemini-embedding-2-preview”,
contents=“Q1 2026 performance report”,
config=types.EmbedContentConfig(output_dimensionality=768)
)
print(f”Dimensions: {len(result.embeddings[0].values)}”)  # 768

Storage impact for a 10-million-document corpus:

Multimodal RAG in practice

The most immediate impact is pipeline simplification. The old pattern required multiple embedding models, custom fusion logic and modality-separated vector spaces. The new pattern is radically cleaner:

from google import genai
from google.genai import types
client = genai.Client()
Index a PDF directly (no manual OCR)
with open(‘financial_report.pdf’, ‘rb’) as f:
pdf_bytes = f.read()
pdf_embedding = client.models.embed_content(
model=‘gemini-embedding-2-preview’,
contents=[types.Part.from_bytes(data=pdf_bytes, mime_type=‘application/pdf’)]
)
Index a meeting recording (no transcription)
with open(‘board_meeting.mp3’, ‘rb’) as f:
audio_bytes = f.read()
audio_embedding = client.models.embed_content(
model=‘gemini-embedding-2-preview’,
contents=[types.Part.from_bytes(data=audio_bytes, mime_type=‘audio/mpeg’)]
)
Query with text — compares against PDF and audio in the same space
query_embedding = client.models.embed_content(
model=‘gemini-embedding-2-preview’,
contents=“What revenue targets were discussed in Q4?”
)
All vectors live in the same space — unified search

The critical point: the text query is compared directly with PDF and audio embeddings — with no intermediate translation layer.

High-value use cases

The combination of native multimodality, an 8,192-token window and 100+ languages opens up cases that were previously infeasible or economically prohibitive:

Where you can use it today

Available via Gemini API (development) and Vertex AI (production with SLA, VPC Service Controls and Vector Search). Documented support across the major libraries:

• LangChain and LlamaIndex — native integration via Google’s embedding class
• Haystack — component available on the hub
• Weaviate, Qdrant, ChromaDB — Google vectorization module
• Vertex AI Vector Search — managed integration with auto-scaling

For Google Cloud teams, the Gemini Embedding 2 + Vertex AI Vector Search + Gemini 2.5 Pro stack delivers a fully managed RAG pipeline with no external dependencies.

Adoption-readiness checklist

Read also

• Gemini 3.1 Pro: AI-Powered Development on GCP
• Dumb Search: 30% of Sales Lost in Retail
• Autonomous Agents: Software That Decides on Its Own