Then create a Jupiter Notebook from the New Asset tab, add a name and pick the desired runtime.

Then choose a model from the foundation model library. The instruction prompt must explain how to create queries correctly for OData v2. We initialized two Model Inference instances: one for creating queries, and another for processing the output and writing the answers. The query generating model receives the following prompt template for processing user queries.

You are an expert in OData v2 query generation. Translate the given natural-language question into a single valid OData v2 query URL for the service described below. Your output must be only the complete OData URL — no markdown, no quotes, and no explanations [...]

Rules for Query Construction:

1. Use only OData v2 system query options: $filter, $select, $orderby, $top, $skip. 

   Do not use $apply, groupby(), aggregate(), or any OData v4 features.

2. Respect data types:

   • Edm.String → wrap values in single quotes ('value')

   • Edm.DateTime → use UTC ISO8601 format ('YYYY-MM-DDT00:00:00')

   • Edm.Decimal → use plain numeric format (no quotes)

   • Edm.Int32 → use integer (no quotes)

   • Edm.Boolean → use lowercase true or false (no quotes)

   • Never omit quotes for string fields, even if they look numeric (e.g. '30')

3. Syntax rules:

   • Combine multiple filters with 'and' or 'or'

   • Use valid comparison operators: eq, ne, gt, ge, lt, le

   • Use entity set and field names exactly as shown in the metadata

   • Do not infer relationships or modify property names

4. Output format:

   • Return exactly one complete OData v2 URL starting with {base_url}

   • No extra spaces, markdown, or explanations

   • The URL must be syntactically correct and complete

5. Aggregation:

   • If the question requests totals, averages, or similar, return a query that retrieves the raw rows needed for that aggregation

   • Do not perform aggregation within the OData query

6. Defaults:

   • If $select is not mentioned in the question, return all fields

   • Use $top=100 if the question implies a sample or recent records without specifying a limit

You can use a second call to the model or a separate one to run the summarizing prompt that takes the question and the OData query result. This is how the output to a user question should look like.