Modeling the AdventureWorks Inventory Database for Azure Search

Modeling structured database content into an efficient search index is rarely a straightforward exercise. Scheduling and change management aside, there exists the challenge of denormalizing source rows away from their table-joined state into search-friendly entities.

This post will explore the AdventureWorks sample data, available online, to highlight common experiences in the transition from database to search. All search references in this post relate to the Azure Search service.

AdventureWorks exposes product information through five related tables:

  1. ProductModel: name
  2. Product: name, color, cost, size, weight, image, category (each row joins to a specific ProductModel)
  3. ProductDescription: description
  4. ProductModelProductDescription: locale (each row joins a ProductModel to a specific ProductDescription for a specific language)
  5. ProductCategory: name, parent category

The naïve approach would be to index all rows from the Product table (joined where appropriate) since the Product table has the most specific information. However, that approach would expose the search index to perceived duplicates in a resultset. For example, the Road-650 model is available in 2 colors and 6 sizes. A query for “road bikes” would then be dominated by 12 instances of the same model, differentiated only by size and color. The other 6 road-specific models would all be relegated to the netherworld of search: page two.

The Road-650 model has 12 options. One-to-many entity rows are best represented as multi-value fields or pre-aggregated-value fields in the search index.

Resolving this issue is not as simple as moving the target index to the ProductModel table. Doing so would ignore the important data in the Product table that should still be represented in search results.

Instead, utilize a search-schema feature that does not have a direct parallel in the database model: Collection(Edm.String). By defining multi-value index fields for ‘color’, ‘size’, and ‘image’ the ancillary information is retained for faceting and filtering without polluting the index with duplicate entries.

Similarly, apply aggregate functions to the numeric Product fields, indexing minListPrice instead of every single product listPrice.

A search for “mountain bikes” now shows discrete bicycle models, while preserving important metadata like color, size, and lowest price.

Unfortunately, this type of modeling cannot be easily achieved through SQL statements alone. Instead, use a simple NodeJS script to load the data and then map it into search-friendly JSON entities.

The final database-search mapping looks like this:

  • model (Edm.String: searchable, filterable, retrievable) from ProductModel.Name
  • description_en (Edm.String: searchable) from ProductDescription for the model where culture=’en’
  • color (Collection(Edm.String): searchable, filterable, facetable, retrievable): unique values from Product.Color for the model
  • size (Collection(Edm.String): searchable, filterable, facetable, retrievable): unique values from Product.Size for the model
  • image (Collection(Edm.String): retrievable): unique values from Product.ThumbnailPhoto for the model
  • minStandardCost (Edm.Double: filterable, facetable, sortable, retrievable): aggregate minimum of all Product.StandardCost for the model
  • minListPrice (Edm.Double: filterable, facetable, sortable, retrievable): aggregate minimum of all Product.ListPrice for the model
  • minWeight (Edm.Double: filterable, facetable, sortable, retrievable): aggregate minimum of all Product.Weight for the model
  • products (Collection(Edm.String): searchable, filterable, retrievable): unique values from Product.Name for the model

After joining the ProductModel table with Product, and ProductDescription, use lodash (or Linq in C#) to quickly transform the resultset:

var records = queryYourDatabase();
var models = _(records)
  .map(function(d) {
    return {
      model: _.first(d).ModelName,
      description: _.first(d).Description,
      colors: _(d).pluck('Color').uniq().compact().value(),
      products: _(d).pluck('ProductName').uniq().compact().value(),
      sizes: _(d).pluck('Size').uniq().compact().value(),
      images: _(d).pluck('ThumbnailPhotoFilename').uniq().compact().value(),
      minStandardCost: _(d).pluck('StandardCost').min(),
      maxStandardCost: _(d).pluck('StandardCost').max(),
      minListPrice: _(d).pluck('ListPrice').min(),
      maxListPrice: _(d).pluck('ListPrice').max(),
      minWeight: _(d).pluck('Weight').min(),
      maxWeight: _(d).pluck('Weight').max(),

The resulting JSON looks like this:

    "model": "HL Road Frame",
    "colors": [
    "products": [
      "HL Road Frame - Black, 58",
      "HL Road Frame - Red, 58",
      "HL Road Frame - Red, 62",
      "HL Road Frame - Red, 44",
      "HL Road Frame - Red, 48",
      "HL Road Frame - Red, 52",
      "HL Road Frame - Red, 56",
      "HL Road Frame - Black, 62",
      "HL Road Frame - Black, 44",
      "HL Road Frame - Black, 48",
      "HL Road Frame - Black, 52"
    "sizes": [
    "images": [
    "minStandardCost": 868.6342,
    "maxStandardCost": 1059.31,
    "minListPrice": 1431.5,
    "maxListPrice": 1431.5,
    "minWeight": 961.61,
    "maxWeight": 1043.26

Finally, here is the SQL query to return the initial recordset. I used the mssql npm module to load the data into my NodeJS app.

  m.Name as ModelName,
  p.Name as ProductName,
  SalesLT.ProductModel m
  SalesLT.ProductModelProductDescription md
  ON m.ProductModelId = md.ProductModelId
  SalesLT.ProductDescription d
  ON md.ProductDescriptionId = d.ProductDescriptionId
  SalesLT.product p
  ON m.ProductModelId = p.ProductModelId

Next: working with multi-level taxonomies in Azure Search.

Comments (2)

  1. anonymouscommenter says:

    Azure Search schemas do not explicitly support multi-level taxonomy categories, but such a feature can

  2. Guillaume says:

    This is all fine for an initial load, but how do you keep the Search index up to date ?
    For example model is renamed or new product is added to the model.
    I guess there is a need for some sort of mapping between the database and the search model ?

Skip to main content