Inference Observability & Feedback Workflow in SAP AI Core

To simplify execution, this tutorial provides a pre-configured Bruno collection containing all required API requests.

This collection includes:

- Inference execution
- Observability APIs
- Feedback APIs
- Label management APIs

👉 Download the Bruno collection from here: Bruno_collections

Import the Bruno Collection

- Open Bruno
- Navigate to Collections
- Click on open Collection
- Upload the downloaded folder files

Configure Environment Variables

After importing the collection:

- Select any request (e.g., Get Token)
- Click on No Environment → Configure
- Provide the following values from your service key:
    - ai_auth_url
    - ai_api_url
    - client_id
    - client_secret
    - resource_group
- Save the environment
- Select the configured environment before executing requests

Generate Access Token - Open the Get Token request - Click Send to generate the access token

Note: If the token expires during execution, regenerate it using the same request.

Before working with inference observability, certain foundational components must be in place.

Setup & Authentication

You must first authenticate your API requests using an access token generated from your SAP AI Core service key. This ensures all subsequent API calls are securely authorized.

Resource & Deployment Setup

Inference execution requires an active deployment. You can use either:

- An orchestration service
- A foundation model deployment

This deployment acts as the endpoint where inference requests are sent and processed.

Object Store (S3) Setup

To store complete inference data (request, response, and feedback), an Amazon S3 object store must be configured.

- Required when using **full persistence mode**
- Not required if storing **metadata only**

Important: Only S3 object stores are supported by inference observability

In this step, you will execute an inference request and enable observability to record the interaction.

📂 Bruno File:

Call orchestration service.yml

Step 1: Open the Request

Navigate to the Call orchestration service request in the Bruno collection. This request is used to send prompts to the deployed orchestration or foundation model.

Step 2: Configure Headers

Ensure the following headers are included:

Authorization: Bearer <access_token>
ai-resource-group: <resource-group>
ai-inference-observability-persistence-mode: full
ai-object-store-secret-name: <object-store-name>

These headers enable inference recording and specify where the data should be stored.

Step 3: Update Request Body

Provide the input prompt in the request body. For example:

{
  "config": {
    "modules": {
      "prompt_templating": {
        "prompt": {
          "template": [
            {
              "role": "user",
              "content": "What is the importance of AI in today's world?"
            }
          ]
        },
        "model": {
          "name": "anthropic--claude-3-haiku",
          "version": "latest",
          "params": {
            "max_completion_tokens": 3000
          }
        }
      }
    }
  }
}

Step 4: Execute the Request

Click Send to execute the request.

Step 5: Observe the Response - The model generates a response - A response header ai-inference-id is returned

This ID uniquely identifies the inference and acts as a reference for all subsequent operations such as retrieval, feedback submission, and labeling.

Explanation

At this stage:

- The inference is executed
- Observability is enabled
- The request and response are recorded

This forms the foundation for tracking and analysis.

Important: Only inferences sent to orchestration services or foundation model deployments are recorded in Inference Observability.

Once an inference is recorded, you can retrieve its details for analysis.

📂 Bruno File

Retrieve one inference.yml

Step 1: Open the Request

Navigate to the Retrieve one inference request.

Step 2: Provide Inference ID

Replace the placeholder with the Inference ID obtained from the previous step.

Step 3: Execute the Request

Click Send to fetch the inference details.

Step 4: Analyze the Response

The response includes:

- Model details
- Input and output tokens
- Latency
- Request and response payload (in full mode)

Explanation

This step allows you to:

- Debug incorrect outputs
- Understand model behavior
- Analyze performance metrics

👉 This is the core of Inference Observability

Retrieve All Inferences

Once multiple inferences are recorded, you can retrieve them collectively to analyze overall usage and system behavior.

📂 Bruno File

Retrieve all inferences.yml

This request returns all recorded inferences within the resource group, enabling broader monitoring and analysis.

Retrieve Inferences Using Labels

You can filter inferences using labels to perform targeted analysis across specific environments or use cases.

📂 Bruno File

Retrieve all inferences with label.yml

This request retrieves inferences that match the specified label criteria.

Feedback helps improve the quality of AI responses over time by capturing user evaluation of model outputs.

📂 Bruno File

Post feedback to an inference.yml

Step 1: Open the Request

Navigate to the feedback request in the Bruno collection.

Step 2: Provide Inference ID

Replace the placeholder with the required inference ID

Step 3: Update Payload

[
  {
    "content": {
      "stars": 5
    }
  }
]

Step 4: Execute the Request

Click Send to submit feedback.

Explanation - Feedback is stored along with the inference - Multiple feedback entries can be added

Important: Feedback is supported only when persistence mode is set to full.

Once feedback is added, you can retrieve it to analyze response quality and user ratings.

📂 Bruno File

Get inference feedback.yml

This request retrieves all feedback associated with a specific inference, enabling evaluation of response quality.

Labels allow you to categorize and organize inference data for better monitoring and targeted analysis.

📂 Bruno File

Post labels to an inference.yml

Step 1: Open the Request

Navigate to the label request.

Step 2: Provide Payload

[
  {
    "key": "ext.ai.sap.com/medium",
    "value": "mobile"
  }
]

Step 3: Execute the Request

Click Send to attach labels.

Explanation

Labels help: - Categorize inferences - Filter results - Perform targeted analysis

Note: - Label keys must use the prefix ext.ai.sap.com - Maximum of 16 labels per inference
- Keys and values must not exceed 64 characters

Cleanup is performed using the delete inference API, where you can specify filters such as time range and labels.

Important: Only metadata is deleted. Request, response, and feedback stored in the object store (S3) are not removed.