Everyone searches for the Copilot Studio chunk size. Here's the honest answer: Microsoft doesn't publish one. Files are chunked and vector-indexed automatically, and you can't tune it. What you CAN control is everything else, and most broken agents are broken by the documented limits nobody reads: the 7 MB SharePoint ceiling, sensitivity labels that silently disable files, and sync delays of 4 to 6 hours. This note collects every published number in one place.
🎯 The Quick Answer on Chunking
If you came here from a search like "copilot studio chunk size tokens", here's the state of it:
| Question | Answer |
|---|---|
| Can I set the chunk size? | No. Chunking is automatic and not configurable |
| Does Microsoft publish the chunk size? | No. No token count, no page count, no character count |
| What IS documented? | Files are "chunked into smaller pieces" and vector-indexed; the orchestrator retrieves the chunks that match the query and the LLM summarizes the top chunks |
| Can I influence it? | Yes, through document structure. See the formatting rules below |
☕ My take: Stop hunting for a magic number. The chunking pipeline is a black box by design, and it changes as Microsoft tunes the platform. The lever you actually have is how you structure your documents, and it's a bigger lever than any token setting would be.
⚙️ How Knowledge Ingestion Actually Works
From the official architecture, the pipeline for uploaded files is:
- Upload: the file lands in the agent's built-in store (Dataverse)
- Chunking: the file is split into smaller pieces for faster processing
- Vector indexing: each chunk is embedded for semantic matching. Indexing time depends on file size
- Retrieval: at question time, the orchestrator pulls the chunks that semantically match the user's query
- Summarization: the LLM writes the answer from the top matching chunks and cites the source
The practical consequence: a chunk has to make sense on its own. If your key fact sits in a table row 14 pages away from the heading that gives it meaning, retrieval can surface the row without the context, and the answer quality shows it.
📏 The Numbers Microsoft Publishes
Agent-level limits
| Limit | Value |
|---|---|
| Knowledge sources per agent | 500 across all types |
| Files uploaded per agent | 500 |
| File upload size | 512 MB per file |
| Instructions length | 8,000 characters |
| Connector payload | 5 MB (450 KB on GCC) |
| Skills per agent | 100 |
| Images in files | Only supported inside PDF files |
File uploads (built-in store)
| Rule | Value |
|---|---|
| Supported types | doc, docx, xls, xlsx, ppt, pptx, pdf |
| Max size | 512 MB per file |
| Password-protected files | Not indexed |
| Sensitivity labels confidential or higher | Not indexed, and they fail silently (see gotchas) |
SharePoint as a knowledge source
This is where most real-world agents quietly break:
| Rule | Value |
|---|---|
| Generative answers file cap without a Microsoft 365 Copilot license | 7 MB per file |
| File cap with M365 Copilot license + tenant graph grounding with semantic search | 200 MB |
| Files referenced by name in a question | Not answerable ("what does file-name.pdf say" fails) |
| Lists per Add knowledge session | 15 |
| List views | Not supported |
| Document libraries (upload path) | Not supported |
OneDrive and SharePoint unstructured sources
| Rule | Value |
|---|---|
| Files per source | 1,000 |
| Folders per source | 50, up to 10 subfolder levels |
| Folder handling | A folder counts as ONE knowledge source including all its content |
| Max file size | 512 MB |
| Sync frequency | Every 4 to 6 hours after ingestion completes |
| Shared folders | SharePoint yes, OneDrive no |
| Page-level citations | PDF only, via the Upload files > SharePoint path; everything else falls back to document-level citations |
| Indexing threshold for page citations | Files under roughly 4 KB fall back to document-level citations |
Authentication is per-user for all unstructured data sources. Users must sign in before the agent can query these sources, and single credential sign-in isn't supported. If your agent works for you but not for a colleague, check their access to the underlying SharePoint or OneDrive content first.
🕳️ The Silent Failures
These are the ones that burn hours, because nothing errors:
- Sensitivity labels fail silently. A file labeled confidential or highly confidential shows as "Ready" in the knowledge list but never contributes to answers. The agent just acts like the file doesn't exist. Same for password-protected documents
- "Ready" doesn't always mean ready. After adding files, the status can show Ready, flip to In Progress, then return to Ready. Only the second Ready means the content is actually indexed and usable. If answers seem to ignore a fresh upload, wait and re-check the status
- XLSX files index, but agents can't run code. Analytical questions ("what's the average of column C") get poor answers. Structured questions need Dataverse or a connector, not a spreadsheet upload
- No glossaries or synonyms. If your company says "colleague" and users type "employee", retrieval takes the hit. Write documents in the vocabulary users actually search with
- ALM doesn't reprocess knowledge. Importing an agent into another environment does not automatically re-ingest knowledge sources. Plan a manual re-add step in your deployment checklist
- Sync is not instant. OneDrive and SharePoint sources refresh every 4 to 6 hours. If someone updates the policy doc at 9 AM, the agent may answer from the old version until mid-afternoon
The 7 MB SharePoint trap is the most expensive one on this list. Without a Microsoft 365 Copilot license in the agent's tenant, generative answers skip any SharePoint file over 7 MB. No error, no warning, just no answer from that file. Split large documents or license the tenant and turn on tenant graph grounding.
📝 Formatting Files So Chunking Works for You
Since you can't tune the chunker, tune the documents. These rules come straight from how the retrieval pipeline works:
| Rule | Why |
|---|---|
| One topic per document | Retrieval matches chunks to queries; focused files produce focused chunks |
| Use a real heading hierarchy | Headings give chunks context and improve both matching and citation quality |
| Make sections self-contained | A chunk that starts "as mentioned above" retrieves badly. Restate the subject in each section |
| Front-load the facts | Put the answer in the first sentences of a section, not after three paragraphs of preamble |
| Split anything over 7 MB | Stays under the SharePoint generative-answers cap and indexes faster |
| Prefer PDF when citations matter | Page-level citations only work for PDFs on the upload path |
| Avoid sprawling multi-page tables | Table rows separated from their headers lose meaning when chunked |
The fastest quality win I've found: take your worst-performing knowledge file and rewrite its section headings as the literal questions users ask. "VPN access for contractors" retrieves better as "How do contractors get VPN access?" because it matches the query embedding almost word for word.
🧭 Which Knowledge Source Type to Use
Uploaded files: stable reference content (policies, manuals, product sheets). Fastest setup, but remember the 4 to 6 hour refresh doesn't apply since files are static until re-uploaded. SharePoint or OneDrive: living documents your team already maintains, accepting the sync delay and per-user auth. Dataverse and enterprise connectors: structured or real-time data. Changes reflect on the next query, access respects each user's security roles, and analytical questions actually work.
How to Scale Your Business with Azure Copilot and Power Platform
My conversation on where Copilot and the Power Platform actually earn their keep in a business, beyond the demo-ware.
🎓 Where This Shows Up Next
Two reasons this cheatsheet is worth bookmarking beyond today's build:
- The AB-620 exam. Microsoft's new AI Agent Builder Associate certification tests planning and configuring agents with knowledge sources as a core domain, and the integration domain (Foundry, MCP servers, connectors) is 40 to 45 percent of the exam. See where certifications fit your path in my Azure Certification Roadmap
- Grounded agents are the pattern now. Whether it's Copilot Studio, Microsoft Foundry, or a custom RAG app, the same discipline applies: source quality and document structure decide answer quality. I walk through the Azure side of this in Building AI Solutions with Azure AI Foundry
📚 Official References
| Resource | What's There |
|---|---|
| Copilot Studio quotas and limits | The authoritative limits table this note is built from |
| Knowledge sources summary | Source types, licensing gates, tenant graph grounding |
| Unstructured data as a knowledge source | OneDrive and SharePoint ingestion details |
| Knowledge in Copilot Studio (Power Platform blog) | The chunking and vector-indexing architecture explained |
Microsoft moves these numbers without much fanfare. If a limit here doesn't match what you're seeing, trust the quotas page, then come tell me so I can update this note. 🔄
