Docling Remote for Dify

A lightweight Dify plugin for a separately deployed Docling Serve. Document conversion, OCR, table extraction and chunking run on your server; the plugin installs no Docling, Torch or model weights.

Русский · Releases · Privacy

Installation

Deploy Docling Serve separately and provision its models and OCR languages. The tested baseline is Serve 1.21.0; see its deployment instructions and the tested server constraints.
Download from release v0.1.4 and install it in Dify → Plugins → Install from local package. Follow your administrator's policy for private plugins.
Configure the provider's Docling Serve URL, such as , without . Set the optional API key matching on Serve.
Run Server health, then add Convert documents to a workflow and bind uploaded files to its Documents input.

The plugin runtime uses Python 3.12. In Docker, refers to the plugin container: connect Serve and the plugin daemon to a shared network, or use an appropriate host gateway. An optional Compose example is provided but has not been integration-tested. Dify Cloud needs an explicitly configured network path to private servers. Use HTTPS across untrusted networks.

Usage and limits

Submit 1–10 files, each nonempty and at most 50 MiB. All inputs are downloaded before conversion starts to avoid expiry during long batches.
Choose Markdown, Docling JSON, HTML, text or DocTags exports. Outputs are , structured and a ZIP under .
OCR languages, standard/VLM pipelines, enrichment and hierarchical/hybrid chunking depend on installed server models. Chunking reuses converted Docling JSON without repeating OCR.
The default PDF backend is : Serve 1.21.0 / Jobkit 1.24.0 rejected the default backend in testing. Select alternatives only with compatible server dependencies.
Timeout is per document: 600 seconds by default, configurable from 10 to 7000. Allow sufficient Dify workflow time for the whole batch. A timeout does not cancel server work; requests are not retried automatically.
Each JSON response and accumulated raw results are bounded to 100 MiB. Mixed batches retain per-document failures; all-failed batches raise an error. remains explicit and may require content review.

Full exports and embedded images stay in the ZIP. Inline JSON removes embedded images and is limited to 2 MiB using Unicode-escaped serialization. Large results retain per-file Markdown when it fits; otherwise marks omitted text and points to the archive. Combined text is limited to 256K characters with an explicit truncation notice. Downstream analysis must check these markers.

examples/workflow_texts.py provides a Dify Code node adapter: bind and to the tool outputs. It returns and , accepts nonempty partial results and rejects failed documents or omitted text. If structured results are unavailable, it rejects truncated combined text.

Compatibility

Previously verified with Dify 1.17.1, native Linux arm64 plugin runtime and Docling Serve 1.21.0: multilingual HTML, scanned PDFs, exports, hierarchical chunking and batch conversion. Conversion success does not guarantee accurate OCR. CUDA, VLM/enrichment inference, Dify Cloud, amd64 runtime and full-scale load have not been verified. Marketplace approval is not claimed.

Development

Python 3.12 and the official Dify CLI are required. Packaging stages an explicit runtime allowlist and writes the package plus into ignored . CI runs tests and checks that matches ; update dependencies together.

For an optional real-server smoke test, set and , then run . An optional points to a local scan containing . Never commit keys, input documents, environments or model weights.

Privacy and support

Documents and filenames go to the administrator-configured Serve instance. Workflow inputs cannot set the server URL. Redirects and environment proxies are disabled; TLS verification is enabled. Raw server error bodies and credentials are not exposed in client diagnostics. Dify and Serve retention policies remain the administrator's responsibility. See PRIVACY.md.

MIT licensed; see LICENSE. Independent integration, not affiliated with Docling or Dify. Report bugs through GitHub Issues without sensitive documents or credentials. Report security issues privately to the repository owner.