[Bug]: DeepSeek custom endpoint returns 400 when uploading PDF files ("file must have a file_id or file_data")

[Learn-Implement]

What happened?

Uploading a PDF file to a conversation using the DeepSeek custom endpoint (deepseek-flash) produces:

400 .messages[1]: file must have a file_id or file_data

This occurs on a brand-new conversation with no prior history. The PDF is small and well under the container's Nginx body limit (client_max_body_size 25M, confirmed inside the container). It is not a request-size issue.

Expected behaviour: LibreChat should either:

1. Use OCR to extract text from the PDF and send it as a text message, or
2. Clearly reject the upload with a user-facing message explaining DeepSeek does not support direct PDF uploads.

Actual behaviour: The PDF is forwarded as a file block in the chat completion request. DeepSeek's Files API only accepts image formats (JPEG, PNG, GIF, WebP) and requires either a file_id (from an upload to https://api.deepseek.com/files) or file_data (base64-encoded image). A PDF cannot satisfy either requirement, so DeepSeek returns 400.

Precedent: Discussion #10988 already documents that DeepSeek requires the OCR workaround for file uploads.

Workaround in use: Switching to Gemini 3.8 Flash for PDF uploads works because Google's API supports PDFs natively.

Version Information

LibreChat v0.8.7
Deployment: Self-hosted Docker Compose (local)
Image: registry.librechat.ai/danny-avila/librechat-dev:latest
Reverse proxy: None (direct port 3080)
Custom endpoints: DeepSeek V4 Pro, DeepSeek Flash, Gemini, Qwen
Redis: redis:7-alpine (docker), USE_REDIS_STREAMS=true

Steps to Reproduce

1. Start a new conversation with the DeepSeek custom endpoint (deepseek-flash selected).

2. Upload any PDF file using the attachment button.

3. Ask a question about the file.

4. Observe the error:
   400 .messages[1]: file must have a file_id or file_data

What browsers are you seeing the problem on?

Firefox, Chrome, Microsoft Edge

Relevant log output

2026-09-23 09:03:13 error: [api/server/controllers/agents/client.js #sendCompletion] Unhandled error type 400 .messages[1]: file must have a file_id or file_data

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[lia-by-librechat[bot]]

Thanks for the report. DeepSeek's documented file input is for images, not PDFs, so this does not look like an Nginx body-size issue. The file_id or file_data error alone does not establish that the PDF's base64 data was missing: LibreChat formats OpenAI-compatible files with nested file.file_data, while DeepSeek documents a different shape. Either way, a PDF should not be sent to DeepSeek as a native file part.

For new uploads on a build with unified file upload (not v0.8.7), route PDFs through text extraction for the custom endpoint:

fileConfig:
  endpoints:
    YourDeepSeekEndpointName: # match the endpoint's name in librechat.yaml
      legacyFileUploadUX: false
      defaultLLMDeliveryPath:
        overrides:
          'application/pdf': text

Please test a new PDF in a new conversation after upgrading. This setting does not repair PDFs already attached under the older routing. If the new test still fails, please provide the exact LibreChat image digest, whether this is a direct chat or saved agent, and a redacted outgoing content-part shape (keys only, no PDF bytes or credentials). Related vLLM/llama.cpp report: #16234.

[Learn-Implement]

Additional test results after updating to rc4 and correcting the config version:

Version: ghcr.io/danny-avila/librechat:v0.8.8-rc4
Digest: ghcr.io/danny-avila/librechat@sha256:929f4491cb9a87d6beb02e9e9ce9a54d4ee05bc56b65b1debe5b1f28a327a4fb

Config version updated to 1.3.17. No "Outdated Config" warning. The DeepSeek block is parsed and visible in the log output:

fileConfig.endpoints.DeepSeek:
fileStrategy: "text"
fallbackToText: true

Test results:

• DOCX on deepseek-flash: works in both standard chat and Agent
• PDF on deepseek-flash in standard chat: fails with "400 .messages[1]: file must have a file_id or file_data"
• PDF on deepseek-flash in Agent: fails with "400 .messages[0]: file must have a file_id or file_data"

Both paths fail. Could you confirm whether:

1. fileStrategy is honoured for custom endpoints in rc4 at all, or
2. An additional config key is required for custom endpoints, or
3. This needs a code change to route PDFs through text extraction for custom endpoints.

[Learn-Implement]

Update: Confirmed root cause and relevant issues

I have tested the fileStrategy: "text" and fallbackToText: true settings in my librechat.yaml (version 1.3.17) across both standard chat and Agent modes. The results are conclusive:

DOCX files work correctly with both deepseek-flash and deepseek-v4-pro.

Images (PNG/JPEG) work correctly with deepseek-flash.

PDF files (both text-based and image-containing) fail consistently with the error: 400 .messages[1]: file must have a file_id or file_data.

This confirms that the fileStrategy setting is not being honoured for custom endpoints in the Agent path in v0.8.8-rc4. The PDF is still being sent as a native file block, which DeepSeek rejects.

This appears to be the same root cause documented in these existing threads:

Discussion #10988 – "DeepSeek and Grok (xAI) configuration: File upload, web search limitations, and required workarounds" – explicitly states that for DeepSeek, "only the OCR approach works".

Issue #13334 – "[Bug]: Agent with DeepSeek model — 'Upload to Provider' sends unsupported file variant" – documents that attachments: false is endpoint-level and is not inherited by agents, causing the same error.

Discussion #6297 – "Inconsistent fileConfig Restrictions for Custom Endpoints" – notes that custom endpoint file restrictions are applied inconsistently between the SidePanel and direct upload.

Could you confirm whether the fileStrategy setting is intended to work for custom endpoints in the Agent path in v0.8.8-rc4, or whether this requires a code change similar to the one proposed in #13334?