[ionut-mihalache-uipath]

Summary

• for models that use the Anthropic Messages API (ChatAnthropic, which includes ChatAnthropicBedrock, and Vertex ChatAnthropicVertex), Analyze Files sends a text file (text/*, application/csv, application/json, application/xml, +json/+xml) as an Anthropic text document (source.type: "text"). The file content is not changed
  • decoded by its UTF-16/UTF-32 byte-order mark if it has one, else as UTF-8 or, failing that, Windows-1252 (Excel's "CSV" encoding)
  • content that decodes to a NUL character is binary (or BOM-less UTF-16) and is sent as a file block, as before, rather than as garbled text
  • the text document carries no guardrail attachment tags (extras): Claude clients send it to the provider as is. File and image blocks keep them
• every other format is still sent as a file block, as before; when the model rejects it, the user now gets a clear error (below)
• a file declared application/octet-stream or with no type whose bytes start with %PDF- is sent as application/pdf, for every provider (SharePoint "Convert to PDF" downloads arrive this way)
• MIME types are compared without their parameters, so application/pdf; name="inv.pdf" is sent as a PDF
• a 400 on a file-analysis request becomes a user FILE_ERROR, "The model could not read the attached files", naming each file and its type, or "The attached files are too large for the model" when the provider reports a context-window overflow
  • files are named with the type they are sent as (a .csv declared application/vnd.ms-excel is named text/csv)
  • when every attachment is a PDF, the error suggests the PDF may be damaged, password-protected or too long instead of "convert to PDF"
  • a 400 carrying the gateway's own (non-blank) detail, a content-filter rejection, and a 400 that names the file limit it hit (too many images, image or document too large) keep the existing mapping
  • the provider's message is still never copied into the error
• the unsupported-MIME-type error now names the attached files

Bedrock Converse, OpenAI and Gemini requests are unchanged; Converse already reads CSV natively.

Why

Analyze Files on Claude failed for a CSV and an unlabeled SharePoint PDF download. The Anthropic Messages API rejects any base64 document whose media type is not application/pdf (document.source.base64.media_type: Input should be 'application/pdf'), even for a correctly typed text/csv, and the user saw a generic "review the model settings" error.

Testing

• unit tests for the text document, decoding, model detection, PDF sniffing, the error mapping and the tool wiring; the text document is pinned through the langchain_anthropic and Vertex Anthropic formatters
• ruff check, ruff format --check, mypy, full pytest
• live on alpha, through build_file_content_blocks_for and the model:
  • anthropic.claude-sonnet-4-5-20250929-v1:0 over Bedrock invoke (UiPathChatAnthropicBedrock) and claude-sonnet-4-6 (UiPathChatAnthropic): UTF-8 CSV, cp1252 CSV, unlabeled PDF and PDF with MIME parameters are all read; an XLSX gets "The model could not read the attached files" naming the file
  • on Bedrock invoke, a CSV labelled application/csv, a UTF-16 CSV with BOM and a JSON file are read; a PNG labelled text/plain is sent as a file and gets the named error
  • the same Bedrock model over Converse: unchanged, CSV and PDFs read

Known limits

• on Claude over the Messages API, Office files (XLSX, DOCX, PPTX) are still not readable; they now fail with an error that names the file and suggests converting to PDF
• the memory node's file attachments are not changed

🤖 Generated with Claude Code

https://claude.ai/code/session_018uaE4Knma5aBySxy4Qxu8n

[Copilot]

🟡 Changes recommended

PDF detection, binary rejection, and PowerPoint ordering contain correctness gaps.

3 open findings

• Check original MIME before normalization for unlabeled PDF detection · New
• Reject binary data before applying the cp1252 fallback · New
• Enumerate PowerPoint slides using presentation relationship order · New

What changed in this PR

Updates Analyze Files to support non-PDF attachments with Claude-compatible models and provide clearer file-related errors.

Changes:

• Converts text and Office Open XML files into Anthropic text documents.
• Detects unlabeled PDFs and Claude model capabilities.
• Improves provider error mapping and attachment diagnostics.

File	Description
pyproject.toml	Bumps package version.
uv.lock	Updates locked project version.
src/​uipath_langchain/​agent/​exceptions/​llm.py	Adds file-analysis error mapping.
src/​uipath_langchain/​agent/​multimodal/​__init__.py	Exports model capability detection.
src/​uipath_langchain/​agent/​multimodal/​invoke.py	Routes attachments by model capability and content.
src/​uipath_langchain/​agent/​multimodal/​text_documents.py	Extracts text from plain and Office files.
src/​uipath_langchain/​agent/​multimodal/​utils.py	Adds PDF sniffing and provider detection.
src/​uipath_langchain/​agent/​tools/​internal_tools/​analyze_files_tool.py	Integrates conversion and file-aware errors.
tests/​agent/​multimodal/​test_text_documents.py	Tests text-document conversion.
tests/​agent/​multimodal/​test_utils.py	Tests model detection and PDF sniffing.
tests/​agent/​test_llm_file_analysis.py	Tests user-facing file errors.
tests/​agent/​tools/​internal_tools/​test_analyze_files_tool.py	Tests Analyze Files integration.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[sonarqubecloud]

Quality Gate passed

Issues
2 New issues
0 Accepted issues

Measures
0 Security Hotspots
97.6% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud