Stringify structured output schema as an attribute

[dinmukhamedm]

Background

It's very useful, when tracing complex apps, to see the structured output schema definition as an attribute. Most major provider SDKs support passing the pydantic (or zod in js) schema to the request, which directs the model to produce the valid JSON and parses the output back to pydantic object upon response. For example:

• OpenAI
• Gemini

Motivation

1. Very useful for debugging
2. Very similar to tool definitions
   • As far as I know OpenTelemetry standards don't support tracing tool definitions, but we do

Proposed solution

Attributes

I'd probably call the attribute something like llm.request.schema or llm.request.structured_output_schema and put a stringified schema value into there.

Implementation

This generally should be as simple as using pydantic.schema_json_of(). https://docs.pydantic.dev/1.10/usage/schema/#getting-schema-of-a-specified-type

[nirga]

Fair point, I'd use gen_ai.request.structured_output_schema to better align with upcoming official otel semantic conventions (we should fix all attributes we have here at some point as well).

[dinmukhamedm]

@nirga just made #3064 for this. Could you please have a look, and if it looks ok to you release that? This will let me add the attribute to the instrumentations