Power Platform Community Forum Thread Details

I am creating an agent in Copilot Studio to summarize PDFs.
The PDF to be read is passed from the trigger "When a new email arrives (V3).
However, it seems that the summary is created using only the content from the first two to three pages, without reading the entire PDF.
If anyone knows the cause and solution, I would appreciate your guidance.

■Instructions
・The PDF file name (including extension) will be passed from the trigger.
・Retrieve the corresponding file (including extension) with the given PDF file name from any of the SharePoints set as the knowledge base, and summarize its content in an easy-to-understand way.
・As preprocessing for the summary, extract all text data from the PDF file.
・Be sure to check until the final page of the PDF and extract all text, not just partial content.
・Summarize based on the extracted text data.
・Post the summarized content on Teams.
Please follow the instructions below for summarization: .......

■Knowledge
SharePoint

■Tool
Teams

■Trigger
When a new email arrives (V3)

{
  "type": "OpenApiConnectionNotification",
  "inputs": {
    "parameters": {
      "subjectFilter": "リム電力レポート",
      "fetchOnlyWithAttachment": true
    },
    "host": {
      "apiId": "/providers/Microsoft.PowerApps/apis/shared_office365",
      "connection": "shared_office365",
      "operationId": "OnNewEmailV3"
    }
  }
}
{
  "type": "InitializeVariable",
  "inputs": {
    "variables": [
      {
        "name": "FileNames",
        "type": "array"
      }
    ]
  },
  "runAfter": {}
}
{
  "type": "Foreach",
  "foreach": "@triggerOutputs()?['body/value']",
  "actions": {
    "Apply_to_each": {
      "type": "Foreach",
      "foreach": "@item()?['attachments']",
      "actions": {
        "Create_file": {
          "type": "OpenApiConnection",
          "inputs": {
            "parameters": {
              "dataset": "https://globaleng0101.sharepoint.com/sites/DX",
              "folderPath": "/Shared Documents/DX推進/Copilotエージェントテスト",
              "name": "@item()?['name']",
              "body": "@item()?['contentBytes']"
            },
            "host": {
              "apiId": "/providers/Microsoft.PowerApps/apis/shared_sharepointonline",
              "connection": "shared_sharepointonline",
              "operationId": "CreateFile"
            }
          },
          "runtimeConfiguration": {
            "contentTransfer": {
              "transferMode": "Chunked"
            }
          }
        },
        "Append_to_array_variable": {
          "type": "AppendToArrayVariable",
          "inputs": {
            "name": "FileNames",
            "value": "@outputs('Create_file')?['body/Name']"
          },
          "runAfter": {
            "Create_file": [
              "Succeeded"
            ]
          }
        }
      }
    }
  },
  "runAfter": {
    "Initialize_variable": [
      "Succeeded"
    ]
  }
}
 
{
  "type": "OpenApiConnection",
  "inputs": {
    "parameters": {
      "Copilot": "crd1d_wzS8f-yR7ZIQw25-ZzROH",
      "body/message": "Use content from @{join(variables('FileNames'),',')}"
    },
    "host": {
      "apiId": "/providers/Microsoft.PowerApps/apis/shared_microsoftcopilotstudio",
      "connection": "shared_microsoftcopilotstudio",
      "operationId": "ExecuteCopilot"
    }
  },
  "runAfter": {
    "For_each": [
      "Succeeded"
    ]
  }
}

Categories:

Calling actions from Copilot Studio

Hey @響遠-04110743-0

A few different things could cause this, and which one it is depends on how your flow and agent are wired, so here are some questions to narrow it down, with what each answer would likely mean:

1. Are you calling the summary agent in the same flow run, right after Create file?
If yes then the file may not be fully indexed yet when the agent runs. SharePoint knowledge indexing happens on a schedule, not on demand, so the agent can be racing the indexer and grounding on an incomplete (or missing) copy. Quick test: run the same file through the agent again an hour later, if it reads more of the document, timing/indexing is a factor.

2. Is the agent reading the PDF through the SharePoint knowledge source, or are you passing the content into the agent's message?
If it's via the knowledge source then that's retrieval (RAG), not a full read. The agent only gets the top few chunks that match the prompt, and for a "summarize" prompt those chunks almost always come from the opening pages , which matches the "only 2-3 pages" symptom exactly. No instruction can override this, because retrieval decides what reaches the model before the instructions run.

3. Is the PDF a digital/text PDF, or a scanned image?
If it's scanned | text extraction returns little or nothing without OCR, so the agent may only pick up whatever thin text layer exists (often just a header or first page).

4. Is the file larger than 7 MB, and does your tenant have a Microsoft 365 Copilot license?
If over 7 MB with no license ( generative answers silently skip the file, with no error).

5. How long is the document?
If it's very long even passing the content directly can exceed the agent's input/token limit, so it may only process the portion that fits.

If 1 or 2 is the cause (my guess, given the symptom), the most reliable fix is to take the index out of the path: extract the text in the flow and pass it into the ExecuteCopilot message directly , like a direct upload rather than a knowledge lookup then instruct the agent to read the full provided document before answering. That isolates the indexing question entirely. Just be aware 3 and 5 can surface once you do (OCR for scans, chunking for long docs).

What does your setup look like on 1 and 2? That'll tell us which path to go down.