By default, the Dataverse 'List rows' action (and many other connectors that return arrays of data) has an implicit limit of 5,000 records per API call to ensure platform stability. If your query matches more than 5,000 records, it will silently return only the first 5,000. To retrieve more, you must explicitly go into the action's Settings and turn on 'Pagination'. You then define a Threshold (e.g., 10,000), which instructs the Power Automate engine to automatically make multiple background API calls (follow next-link tokens) until the threshold is met or all records are retrieved.

AI Builder Document Processing models have strict technical limits to ensure performance and reliability. The maximum allowed file size for a document sent to the extraction API is 20 MB. Furthermore, if a multi-page document is submitted, the model will only process the first 500 pages. Files exceeding the 20 MB limit or formats that are unsupported will cause the Power Automate action to fail.

The 'Fallback' system topic is triggered whenever the agent cannot determine the user's intent, cannot find a matching custom topic, and cannot generate an answer from its designated knowledge sources. By customizing the Fallback topic, you can add conditional logic to track how many times the user has hit the fallback (using a counter variable), provide the custom message ('I only handle IT queries'), and then use the 'Transfer to agent' node after the counter reaches two.

The standard, Microsoft-recommended pattern for handling action timeouts in Power Automate is using the ISO 8601 duration format in the action's Timeout settings combined with 'Configure run after'. By setting the timeout to 'PT3D' (Period Time 3 Days), the action will intentionally fail with a 'TimedOut' status if no response is received. You then configure the subsequent action (the Dataverse update) to execute specifically when the previous action 'has timed out' (and typically uncheck 'is successful'). Parallel delays cause messy execution states, and 'Escalation' properties require specific premium or custom handling not native to standard approvals.

A core principle of Responsible AI and intelligent application design is keeping 'humans in the loop'. Because AI models like the Business Card Reader can misread text (OCR errors), you should never blindly patch the output directly to a database. The best practice is to map the AI control's outputs to the `Default` property of standard Text Input fields on the screen. The user can then visually verify the extracted data, type in any corrections, and click a separate 'Save' button that patches the values from the Text Input controls, not directly from the AI control.

For the native Copilot control in Canvas apps to query data and provide intelligent answers based on a specific Dataverse table, that table and its relevant columns must be indexed by Dataverse Search. Copilot relies on the Dataverse Search infrastructure to quickly retrieve context and generate accurate natural language responses. If the table is not enabled for Dataverse Search, it will not be available as a selectable data source for the Copilot control.

The AI Hub prompt designer provides a built-in testing experience. When you define input variables (like 'CustomerComplaint'), a 'Sample data' field appears for each variable in the right-hand testing pane. Makers can input test values into these fields and click 'Test prompt' to evaluate the model's response directly within the designer, allowing for rapid iteration before saving or publishing the prompt.

A '429 Too Many Requests' error is a transient error indicating rate limiting. The Microsoft best practice for handling transient HTTP errors is to configure the built-in 'Retry Policy' within the action's settings. By setting an Exponential or Fixed interval retry policy, the Power Automate engine will automatically pause and re-attempt the HTTP call a specified number of times before finally marking the action as failed. This avoids building messy 'Do until' loops with delay variables.

Generating net-new, creative text (like a blog post) based on a few keyword inputs is the primary use case for Large Language Models (LLMs) like GPT. In AI Hub, this is achieved by creating a custom Prompt that utilizes the Azure OpenAI GPT models. Text Recognition (OCR) reads text from images. Category Classification sorts text into buckets. Key Phrase Extraction pulls main points from existing text. None of these generate new, long-form content.

While GPT models (custom prompts) are excellent for text generation and unstructured text summarization, they are not natively designed to process image-based/scanned PDFs to extract highly structured key-value pairs at scale. The 'Document Processing' model in AI Builder is purpose-built for extracting structured data (like specific fields from shipping manifests, invoices, or forms) from PDFs and images.