`MarkItDownMcpReader` ignores tool IsError (returns error text as content)`

[mldisibio]

Description

MarkItDownMcpReader (Microsoft.Extensions.DataIngestion.MarkItDown)

1. Failed conversions are silently returned as document content. When the tool reports a failure, it returns a CallToolResult with IsError = true and the error message in Content. ConvertToMarkdownAsync returns the first text block without checking IsError, so the error string becomes the "markdown" body of the document. Downstream this produces a single bogus chunk and the ingestion is reported as successful.

Reproduction Steps

Using Microsoft.Extensions.DataIngestion.MarkItDown 10.7.0-preview.1.26309.5 against a MarkItDown MCP server (mcp/markitdown, --http):

  var reader = new MarkItDownMcpReader(new Uri("http://localhost:3001/mcp"));

  // Any input the server cannot convert makes the tool return IsError=true with the error in Content.
  // Simplest reproducible case: a path the server cannot resolve.
  var doc = await reader.ReadAsync(new FileInfo("does-not-exist-or-unconvertible.docx"), ct);

  // No exception is thrown. doc has a single text element whose value is the tool's error string, e.g.:
  //   "Error executing tool convert_to_markdown: [Errno 2] No such file or directory: '...'"
  // (verified directly: CallToolResult.IsError == true, and the message is carried in Content as text)

Expected behavior

When the tool returns IsError = true, the reader should throw (or otherwise surface a failure) instead of returning the error text as document content.

Actual behavior

The error message string is returned as the document body; no exception; the failure is invisible to callers and logs.

Relevant current code (MarkItDownMcpReader.ConvertToMarkdownAsync):

Dictionary<string, object?> parameters = new() { ["uri"] = dataContent.Uri }; // always a data: URI
var result = await client.CallToolAsync("convert_to_markdown", parameters, cancellationToken);
// no check of result.IsError before returning the first text block
foreach (var content in result.Content)
    if (content.Type == "text" && content is TextContentBlock textBlock)
        return textBlock.Text;

Regression?

Not a regression — Microsoft.Extensions.DataIngestion.MarkItDown is preview (10.7.0-preview.1.26309.5); behavior present since the reader was introduced.

Known Workarounds

Bypass MarkItDownMcpReader for the conversion call: use ModelContextProtocol directly to invoke convert_to_markdown, check CallToolResult.IsError, then round-trip the returned Markdown back through MarkItDownMcpReader's stream overload so the internal MarkdownParser still builds the IngestionDocument. Works, but reimplements the reader's core.

Configuration

• .NET 9 (SDK 9.0.315), Windows 11 x64.
• Microsoft.Extensions.DataIngestion.MarkItDown 10.7.0-preview.1.26309.5, Microsoft.Extensions.AI 10.7.0, ModelContextProtocol.Core 1.2.0.
• MarkItDown MCP server: mcp/markitdown 1.8.1, --http (Streamable HTTP).
• Not configuration-specific.

Other information

Proposed fixes:

• IsError (pure bug, no API change): in ConvertToMarkdownAsync, if result.IsError == true, throw an InvalidOperationException containing the returned text.

[aaleemsolutions]

Hi, I'd like to work on this issue. I will implement the proposed fix to check result.IsError and throw an InvalidOperationException in ConvertToMarkdownAsync.