By adding a comment like // what does function F do? to the prompt, you are instructing the GPT model to provide an explanation of the function's behavior. This approach directly asks the model to explain the code, which will help generate an explanation without requiring changes to the model or additional resource-intensive configurations.

To identify whether model input, model output, or expanded tool usage is driving the cost spike, you should focus on token usage.
Granular tracking of token usage breaks down the costs directly:
Input vs. Output Tokens: Tracing token usage isolates whether the user prompt (input) has grown or if the model is generating longer, more expensive responses (output).
Tool Calling: Expanded tool usage consumes extra tokens each time the agent schema processes an execution loop, which will clearly reflect as inflated token counts.
Reference:
https://www.mindstudio.ai/blog/subtraction-principle-agent-harness-optimization

The best option in this scenario is a custom Azure Content Understanding in Foundry Tools analyzer that defines the required fields as the extracted fields and the returned confidence scores for routing.
Custom Field Targeting with Confidence Scores: Azure Content Understanding allows you to build a document analyzer with a user-defined schema. By defining your required fields (InvoiceNumber, InvoiceDate, VendorName, and TotalAmount), the service will handle extraction across varying layouts and return a dedicated field-level confidence score. You can natively evaluate these confidence metrics to implement your supervisor routing workflow.
RAG-Ready Output Structure: The service naturally returns highly structured JSON payloads.
This makes it perfectly optimized to be ingested directly as grounding data for a Retrieval- Augmented Generation (RAG) agent.
Incorrect:
[Not C]
The prebuilt-layout analyzer extracts raw structural elements such as blocks of text, hierarchy, selection marks, and complete tables. It does not automatically classify or cleanly isolate specific target entities (like vendor name or total amount) into dedicated schema properties, leaving you with heavy post-processing work to isolate the text.
[Not D]
The prebuilt-documentSearch analyzer is optimized for broad, layout-aware text extraction for indexing. Relying on Azure AI Search's search.score for supervisor routing is conceptually flawed; search.score represents a relevance score for a search query ranking rather than an accuracy metric for data extraction confidence.
Reference:
https://learn.microsoft.com/en-us/azure/ai-services/content-understanding/document/analyzer-improvement

Agentic retrieval in Azure AI Search uses a large language model to decompose a complex query into focused subqueries, runs them in parallel, semantically reranks each set of results, and merges them into a unified response, while taking conversation history into account. That is precisely the behaviour the scenario describes.

Correct:
* You add a reflection pass that regenerates the response if the required clauses are missing.
This is Self-Correction Strategy: A reflection pass allows an agent to evaluate its own initial output against specified constraints (e.g., checking for the presence of mandatory regulatory clauses). If the required text is missing, the agent triggers a programmatic self-correction or regeneration loop to include them before final delivery.
Incorrect:
* You increase the value of the max_tokens parameter.
Increasing the max_tokens parameter prevents the response from being cut off mid-sentence due to length constraints. However, it does not force the model's logic to explicitly include missing information that it chose to leave out earlier in the text.
* You increase the value of the temperature parameter.
Raising the temperature parameter increases randomness and creativity. For rigid compliance tasks like summarizing regulatory documents, higher temperature actually increases the risk of hallucination and omission.
* You run an evaluation flow that scores responses for completeness and blocks responses that fall below a defined threshold.
Evaluation Flow Block: Running an evaluation flow to score completeness and blocking bad responses identifies and stops low-quality outputs, but it does not fix or actively improve the response completeness. It simply filters failures out of the system.
Reference:
https://pub.towardsai.net/reflection-with-llm-how-to-make-ai-review-its-own-work-2db122fca1d8