How to Automate Supply Chain Risk Reports: A Guide for Developers
Do you use Python? If so, this guide will help you automate supply chain risk reports using AI Chat GPT and our News API.
A news agent can pick the wrong search tool before it has seen any parameters. If all tools are described as “search for relevant information”, a detailed schema may arrive too late to assist the model in distinguishing between news discovery and an internal document lookup.
On November 19, 2026, Google launched Tool Retrieval in preview in the Gemini Interactions API. According to Google’s announcement, deferred tools retain their names and descriptions in the model’s context while their parameter schemas load on demand. Test that short description apart from the complete schema.
For a news application, the first release question is whether the agent selects the correct operation from that short view. A well-formed request to the wrong data source can yield a compelling answer to a different question.
Google’s approach uses tool_search and the defer_loading flag to pull schemas on demand. An example Google provides shows the tool search, returned schema and function call in interaction.steps. These records enable a programmer to examine selection and execution of a function separate from one another.
Practically speaking, the difference is between identifying a function and identifying news. Tool Retrieval selects a function. Once a given function has been selected, a news provider then retrieves news content.
Describe the purpose of a tool in the tool’s visible description. Document parameters with the type and limits, but don’t leave the tool’s purpose only in the parameter documentation.
Let’s look at a hypothetical example. A research assistant needs to find reporting published this week on port delays for European automakers. The registry includes a news search wrapper and an internal document search wrapper. Both descriptions say, “Search relevant information.”
Let’s say the assistant selects the internal search option and pulls a planning memo on last year’s delays. The call was successful and the memo contains the relevant industry information. The mistake was source choice, not an invalid parameter or no result. A fluent summary of the memo would still not satisfy the request for recent reporting.
Rename and describe the wrappers so that the distinction is visible before either of the schemas load. Let’s say for the news wrapper, we name it search_news_reporting and describe it as, “Find published news articles and relevant passages for a natural language research question. Use for current reporting. Apply the requested publication date range. Returns source-linked search results, not a completed answer.”
Let’s say for the internal document wrapper, we name it internal_document_search, and describe it as, “Search the organization’s internal documents. Use for plans and records; this does not search published news.” These are application-owned wrapper examples, not prescribed Webz.io function names.
The descriptions give a general idea about what each operation can and cannot do. They don’t need to list every possible question. Like the news wrapper description, avoid saying that an operation retrieves all the information or verifies the truth of the information reported unless that’s actually the case.
Run the port-delay question against the revised registry, and see which schema is requested. The presumed choice is the news wrapper. Next, examine the executed call: did its arguments maintain the requested publication window, and did the returned content come from the specified source type?
Webz.io’s News Search API is a suitable option in this case. It accepts natural-language questions and returns passages, as well as related articles. It supports filters like publication date and language. The API fits the discovery step in this example. The wrapper should translate its offered controls to the actual API request. A description claiming it supports a date restriction is not sufficient if the adapter drops it.
Use the companion request “Find our internal plan for handling those port delays” as a contrast case. The expected choice is now internal search. If both requests always select news, the new description has just made one tool more attractive; it has not established the intended distinction.
Make your expected choices before running your cases. Keep the registry version and the chosen operation with any failures. If the selection is correct, but the date argument is missing, check the construction of the argument. If the wrong schema is requested, look into the visible name and description. Keeping those failures separate makes the next repair more specific.
An assistant that solely provides news may use its news search function for almost all the tasks. Pre-loading that schema is justifiable. An assistant that combines news with a few occasional business tools has a good case for considering deferred loading.
Google allows frequently used tools to stay loaded while others are deferred. This is a configurable option and is not evidence of a savings case for your application. Compare equivalent tasks, check selection failures, input tokens and elapsed time before deciding.
Webz.io’s News API versus News Search API comparison explains the core retrieval difference. Once that difference is understood, check to see if the agent can recognize the wrapper that exposes it. Ship the port-delay case only if the news request selects news and the internal request selects internal records, and if the executed arguments preserve the user’s scope.
Do you use Python? If so, this guide will help you automate supply chain risk reports using AI Chat GPT and our News API.
Use this guide to learn how to easily automate supply chain risk reports with Chat GPT and news data.
A quick guide for developers to automate mergers and acquisitions reports with Python and AI. Learn to fetch data, analyze content, and generate reports automatically.