November 17th, 2026
0 reactions

Simplify AI retrieval with embedding generation using Microsoft-Embedding-1 

Product Manager

Most enterprise knowledge doesn’t live in clean text. It lives in charts, tables, scanned pages, and diagrams, where text-only search often misses it: 

  • Financial reports: An analyst asks, “Which region’s revenue fell in Q3?” The answer is a bar in a chart on page 14, and no sentence on the page says it. 
  • Technical manuals: A field technician asks, “What torque does the intake valve bolt need?” The spec is a callout on an exploded-view diagram in a scanned PDF. 
  • Labor reports: An analyst asks, “Which EU nationalities together make up almost 30% of all posted workers?” The answer is the sum of a few slices of a pie chart. 
  • Product catalogs: A shopper searches for “something to wear on a rainy hike.” The listing says “lightweight waterproof jacket,” and the photo shows a hooded shell on a muddy trail. 

Introducing Microsoft Embedding 1

Today we’re introducing Microsoft Embedding 1, a state-of-the-art multimodal embedding model developed by Azure Cosmos DB and Microsoft Research. It embeds text, images, and full document pages into one vector space, so a single text query can find the passage, photo, or chart that answers it. The model is available through Azure Data AI today and is coming soon to Microsoft Foundry.

Simplify AI retrieval with embedding generation using Microsoft-Embedding-1 

State-of-the-art retrieval on visual documents 

ViDoRe v3 part of the MTEB leaderboard, is a benchmark for retrieval over visually rich enterprise documents such as reports and manuals. On ViDoRe v3, Microsoft Embedding 1 ranks #1 in its size class with a score of 66.24 nDCG@10, as of November 16, 2026.

Microsoft Embedding 1 (microsoft-embedding-1-large) is trained to output both a dense retrieval model and a late-interaction model. You can request one dense vector per input, a multi-vector representation, or both. Both representations score ahead of Cohere Embed v5 Pro, Gemma Embedding 2, voyage-multimodal-3.5, and others on ViDoRe v3. 

Late-interaction models are increasingly used for precise retrieval over complex content, where text-only dense retrieval struggles most. They emit one vector per input token and score a pair of inputs (text, image, or document page) with MaxSim over their vectors. On ViDoRe v3, multi-vector representations add nearly 6 points of nDCG@10 over dense output.  

Storing one vector per token can be costly, though. For a page with about 1,250 visual tokens, a late-interaction model can output 1,250 vectors × 1,024 dimensions, which is about 2.5 MB. To address this, we trained the model so you can control the output size by limiting:  

  1. The number of vectors per input. 
  2. The dimensions per vector. 

For example, you can request 32 vectors with 128 dimensions each in float16, which comes to 8 KB per page. That’s in the same ballpark as a single dense vector. The model still maintains strong relevance, giving up about 3.4 points of nDCG@10 for a roughly 300-fold reduction in storage. 

Text retrieval 

On text retrieval, Microsoft Embedding 1 scores on-par with OpenAI text-embedding-3-large on the BEIR benchmark (15 tasks, nDCG@10), with vectors about 17% smaller (2,560 vs. 3,072 dimensions). At 1,536 dimensions, the same size as text-embedding-3-small, it scores 3.8 points higher. 

Model  Dimensions  BEIR average 
Microsoft Embedding 1  1,536  54.92 
Microsoft Embedding 1  2,560  55.29 
OpenAI text-embedding-3-small  1,536  51.08 
OpenAI text-embedding-3-large  3,072  55.43 

Microsoft Embedding 1’s largest gains over text-embedding-3-large are on ClimateFEVER (+6.3), HotpotQA (+3.6), Natural Questions (+3.3), DBPedia (+3.0), and MS MARCO (+1.8). text-embedding-3-large leads most on TREC-COVID. Microsoft Embedding 1 scores are measured with fp16 vectors, as served.  

Choose your balance of accuracy and storage 

Two settings let you trade accuracy against storage:  

  • Multi-vector output. Choose how many vectors the model returns per page or passage. A reasonable starting point is 32 vectors per input with 128 dimensions each. On ViDoRe v3, even this setting beats the best dense result by 2.5 points.  
  • Flexible dimensions. With Matryoshka representation learning (MRL), you can pick a dense vector size of 1,024, 1,536, 2,048, or 2,560 dimensions. Smaller vectors cost less to store and are slightly faster to search.  

Line chart of ViDoRe v3 retrieval quality (nDCG@10) versus average storage per document on a log scale. Dense scores 57.08 to 60.37 at 2 to 5 KB. Multi-vector capped at 32 vectors scores 62.88 to 64.18 at 8 to 64 KB. Unconstrained multi-vector scores 65.63 to 66.24 at about 300 to 2,500 KB. Figure: ViDoRe v3 retrieval quality (nDCG@10) versus average raw FP16 vector storage per document. Each point is a different vector dimension.

The chart above shows how each configuration trades retrieval quality for storage on ViDoRe v3. Capping multi-vector output at 32 vectors keeps most of the quality gain over dense vectors at a fraction of the storage of unconstrained multi-vector output.  

For text-heavy content, dense vectors work well. For visual documents such as reports and manuals, multi-vector output gives the highest retrieval quality. Multi-vector output requires a retrieval system that supports MaxSim scoring.  

Generate embeddings in three steps 

You can use Microsoft Embedding 1 without deploying or hosting a model: 

  1. Enable embedding generation. Today, you enable it on an Azure Cosmos DB account, and inference is provisioned for you.
  2. Generate embeddings by calling the endpoint with the Azure Data AI client library (azure-data-ai) for Python or .NET. 
  3. Store and search the vectors in your vector store, such as Azure Cosmos DB for dense vectors. 

Simplify AI retrieval with embedding generation using Microsoft-Embedding-1 

Install the Azure Identity library and the preview of the Azure Data AI library:

pip install azure-identity
pip install "azure-data-ai==1.1.0b1"

This example embeds a search question and a page from a scanned equipment manual. Set your endpoint, image URL, and SAS token as environment variables.

import os

from azure.data.ai import InferenceClient
from azure.identity import DefaultAzureCredential

with DefaultAzureCredential() as credential:
    with InferenceClient(
        endpoint=os.environ["EMBEDDING_ENDPOINT"],
        credential=credential,
    ) as client:
        response = client.get_embeddings(
            model="microsoft-embedding-1-large",
            inputs=[
                {
                    "kind": "query",
                    "content": "What torque does the intake valve bolt need?",
                },
                {
                    "kind": "image",
                    "url": os.environ["IMAGE_BLOB_URL"],
                    "sasToken": os.environ["IMAGE_SAS_TOKEN"],
                },
            ],
        )

# Results are returned in input order.
embeddings = [item.embedding for item in response.data]

Install the preview of the Azure Data AI library and the Azure Identity library:

dotnet add package Azure.Data.AI --prerelease
dotnet add package Azure.Identity

This example embeds a search question and a page from a scanned equipment manual. Set your endpoint, image URL, and SAS token as environment variables.

using Azure;
using Azure.Data.AI;
using Azure.Identity;

var client = new InferenceClient(
    new Uri(Environment.GetEnvironmentVariable("EMBEDDING_ENDPOINT")!),
    new DefaultAzureCredential());

Response<EmbeddingResult> response = client.GetEmbeddings(
    model: "microsoft-embedding-1-large",
    inputs: new[]
    {
        new EmbeddingsInputContent
        {
            Kind = EmbeddingContentKind.Query,
            Content = "What torque does the intake valve bolt need?"
        },
        new EmbeddingsInputContent
        {
            Kind = EmbeddingContentKind.Image,
            Uri = new Uri(Environment.GetEnvironmentVariable("IMAGE_BLOB_URL")!),
            SasToken = Environment.GetEnvironmentVariable("IMAGE_SAS_TOKEN")
        }
    });

// Results are returned in input order.
var embeddings = response.Value.Data
    .Cast<MultimodalDenseEmbedding>()
    .Select(item => item.Embedding.ToObjectFromJson<float[]>())
    .ToArray();

Turn embeddings into relevant results 

Store each embedding with its content and metadata, embed the user’s question with kind: “query”, and retrieve the closest items. In Azure Cosmos DB, store dense vectors in a container with a matching vector policy and index, and query them with `VectorDistance`. Show the results directly, or pass them to an AI model as grounding context. 

 Pay for what you use 

We charge for the input tokens you process, so you can start with a small dataset and expand as your application needs grow. 

Input  Price per 1 million input tokens 
Text  USD $0.12  
Images  USD $0.40  

Conclusion

Charts, tables, and scanned pages hold some of your most important answers, and text-only search can’t see them. Microsoft Embedding 1 can, with the highest published ViDoRe v3 score for a single retrieval model and no model to deploy or host. Pick the balance of accuracy and storage that fits your app, and pay only for what you embed. Your best answers are already in your documents. Now your search can find them.

 

About Azure Cosmos DB

Azure Cosmos DB is a fully managed NoSQL and vector database for modern applications, including AI apps. With SLA-backed performance and availability, automatic scaling, and global distribution, it helps developers build fast, reliable applications.

Find more developer resources at the Azure Cosmos DB developer hub.

Stay connected with Azure Cosmos DB on X, YouTube, and LinkedIn.

Category

Author

Khelan Modi
Product Manager

Hey there! I'm Khelan Modi, Product Manager at Azure Cosmos DB. My mission is to build a better MongoDB service for our cosmic universe! Join me on this cosmic journey from Redmond, and together, we'll shoot for the stars! 🌠 Let's connect and explore the wonders of cloud tech!

0 comments