Building and customizing solutions using Microsoft 365 Copilot APIs and tools
Hello Kalla, Jahnavi ,
The root cause is that Microsoft Copilot Studio knowledge grounding does not support structured formats like JSON, TXT, or CSV, and only indexes document-style sources such as PDFs, Word, PowerPoint, Excel, web URLs, and SharePoint pages. Because JSON lacks semantic layout, hierarchy, and chunkable text structure, the ingestion pipeline skips citation metadata generation, which results in citations that work during local testing but become non-clickable after publishing since no tenant-hosted URI or Graph resource ID exists. Clickable citations require a supported document type hosted in SharePoint or OneDrive with indexable layout and Graph-mapped chunks, which JSON fails to provide.
Recommended production workarounds include converting JSON into document-style content (PDF/Word/SharePoint pages), using Azure AI Search for structured data retrieval via plugins, or auto-generating SharePoint pages from JSON. This limitation persists across environments and is expected to remain, as the platform currently focuses on unstructured document grounding rather than structured file ingestion.