What this covers

This exercise (about 30 minutes) puts the vector search module into practice. You deploy an Azure Cosmos DB for NoSQL account with vector search enabled, create a container with a vector embedding policy and vector index, load support tickets that already include embeddings, and find similar tickets with VectorDistance from Python, tested through a Flask web app. The scenario: help support agents resolve a new case faster by surfacing similar past cases.

Exam weight: Develop AI solutions by using Azure data management services (30–35%)

Skills measured

"Store and retrieve embeddings and execute vector similarity search for semantic retrieval"

Key concepts

Lab tasks

Download the starter files and configure the deployment script.
Deploy a Cosmos DB for NoSQL account with the vector search capability.
Write Python functions for vector similarity search.
Create a container with vector embedding and indexing policies.
Test search through the Flask app.

Prerequisites: Azure subscription, Visual Studio Code, latest Azure CLI, Python 3.12+.

Pre-computed embeddings

The sample tickets ship with vectors already generated, so the lab doesn't need an embedding deployment to load data. The rule still applies: the query vector must come from the same model and dimension count as the stored vectors, or similarity scores are meaningless.

Container design for the scenario

Partition key such as /category so agents can search within a product area on one partition.
Vector policy: /embedding, float32, cosine, dimensions equal to the sample model's output.
Index policy: exclude /embedding/* from range indexing; add a vector index. With fewer than 1,000 vectors, quantizedFlat and diskANN fall back to a full scan, so a small lab dataset behaves like brute force anyway.

Search function contract

Input: query vector and optional filters. Output: top N tickets with ID, title, resolution and similarity score. Return the score so the UI can hide weak matches.

How it works

Hands-on

Azure portal

Open the deployed account > Features and confirm Vector Search for NoSQL API is on.
In Data Explorer, open the container Settings and review Vector Policy and Indexing Policy.
Run a vector query and note the RU charge in Query Stats.

Azure CLI

Azure CLI / Bash

az cosmosdb create --name <account-name> --resource-group <resource-group> \
  --locations regionName=eastus \
  --capabilities EnableNoSQLVectorSearch

az cosmosdb sql database create --account-name <account-name> \
  --resource-group <resource-group> --name supportdb

Python

from azure.cosmos import CosmosClient, PartitionKey
from azure.identity import DefaultAzureCredential

client = CosmosClient("https://<account-name>.documents.azure.com:443/", DefaultAzureCredential())
db = client.get_database_client("supportdb")

tickets = db.create_container_if_not_exists(
    id="tickets",
    partition_key=PartitionKey(path="/category"),
    vector_embedding_policy={"vectorEmbeddings": [
        {"path": "/embedding", "dataType": "float32", "distanceFunction": "cosine", "dimensions": 1536}]},
    indexing_policy={
        "includedPaths": [{"path": "/*"}],
        "excludedPaths": [{"path": "/\"_etag\"/?"}, {"path": "/embedding/*"}],
        "vectorIndexes": [{"path": "/embedding", "type": "quantizedFlat"}],
    },
)

def find_similar(query_vector: list[float], category: str | None = None, top: int = 5):
    where = "WHERE c.category = @category" if category else ""
    query = f"""
        SELECT TOP @top c.id, c.title, c.resolution,
               VectorDistance(c.embedding, @v) AS score
        FROM c {where}
        ORDER BY VectorDistance(c.embedding, @v)"""
    params = [{"name": "@v", "value": query_vector}, {"name": "@top", "value": top}]
    if category:
        params.append({"name": "@category", "value": category})
        return list(tickets.query_items(query, parameters=params, partition_key=category))
    return list(tickets.query_items(query, parameters=params, enable_cross_partition_query=True))

Exam traps

Lab decision	Why
Enable `EnableNoSQLVectorSearch` before creating the container	Vector policies are rejected otherwise
Set vector policy at creation	It can't be edited in place later
`TOP N` in every vector query	Bounds RU cost and latency
Optional category filter with `partition_key=`	Single-partition, cheaper search
Return the score	UI can apply a relevance threshold

Quick review

The lab takes about 30 minutes and uses Python 3.12+ and the Azure CLI.
Vector search is enabled at account level first.
The container gets both a vector embedding policy and a vector index.
Tickets arrive with pre-computed embeddings.
Queries use VectorDistance with TOP N and ORDER BY.
Filters plus partition_key= keep searches on one partition.

Practice

Practice: Cosmos DB vector search questions

Exercise - Build a semantic search application with Azure Cosmos DB for NoSQL