5.2.14. Docling (Document Processor engine)

Docling is the engine behind the Document Processor role, selected with DOC_PROCESSOR_SOURCE. See Document Processor for setup, configuration and integration.

1. Engine quick reference

• Image (GPU): pytorch/pytorch:2.13.0-cuda12.6-cudnn9-runtime (digest-pinned; used as BASE_IMAGE in the GPU provider Dockerfile); the provider requirements keep torch==2.13.0 and its matching torchvision==0.28.0 patch pair.
• License: MIT (IBM)
• Activation: DOC_PROCESSOR_SOURCE=docling-container-gpu (or docling-localhost for host-installed Docling)
• In-container port: 8000
• Host port: ${DOC_PROCESSOR_PORT} (computed from BASE_PORT by the bootstrapper)
• Readiness: GET /health starts configured converter construction off the API event loop and returns 503 starting until it succeeds. Invalid pipeline or device configuration returns 503 unavailable; health reports the converter only and does not claim lazily loaded model artifacts.

JSON image references. Uploaded Docling JSON keeps embedded images. Both providers ignore filesystem paths, file: URIs and unsupported image schemes before enrichment or export. Only data:, http: and https: references remain. The log reports their discarded count without exposing paths. Images from other input formats keep their existing handling.

Timeouts and fallback.

• Backend and Celery extraction wait up to DOCLING_INFERENCE_TIMEOUT_SECONDS plus 30 s for Docling. TIKA_TIMEOUT_SECONDS does not apply.
• Backend POST /documents/extract routes long-tail formats straight to Tika. These include legacy Office (.doc, .xls, .ppt), .epub, mail, RTF, OpenDocument and archives. Atlas's Docling providers answer an unsupported format with 500, not 415.
• Celery RAG ingestion follows the profile's parser_order. It can try Tika after Docling fails.
• Kong cuts api.localhost requests at 300 s. The HTTP client gets a 504 while the conversion continues.
• A Celery RAG ingestion has its own limits, RAG_INGESTION_TASK_SOFT_TIME_LIMIT_SECONDS and RAG_INGESTION_TASK_TIME_LIMIT_SECONDS. By default they are the larger of 3840 / 3900 s and the global Celery limits.

2. Capabilities & limitations

Support tier: experimental — Capability contract declared; no cited cold-start, workflow, or upgrade qualification run yet (evidence at v0.1.0).

Capability	Status	Verification	Notes
Docling document conversion sources	partial	tested	Atlas provides an NVIDIA GPU container and an existing-host endpoint, but no CPU container or Atlas-managed native Docling lifecycle.
Structured extraction and bounded chunking	supported	tested	The provider converts documents once and renders structured markdown or JSON with validated OCR, table, formula, code, chunk-size, overlap, and total-chunk controls.
Authenticated bounded provider API	partial	tested	Atlas-managed Docling routes require a generated bearer token and enforce upload, admission, and inference deadlines by default, but AUTH_MODE=disabled is an explicit rollback.
Truthful model readiness	partial	tested	Health stays unavailable until converter construction succeeds, but it does not certify every lazily loaded model artifact needed by a later document.
LightRAG conversion bundle	supported	tested	An authenticated internal route renders the JSON and Markdown bundle consumed by the isolated asynchronous LightRAG compatibility adapter.

3. Dependencies & Integrations

3.1. Current — Upstream (this service calls)

No upstream calls.

3.2. Current — Downstream (services that call this)

Service	Category	Status
kong	infra	current
docling-lightrag-adapter	media	optional: DOC_PROCESSOR_SOURCE=docling-container-gpu or docling-localhost
celery	agents	current
n8n	agents	current
backend	apps	current
jupyterhub	apps	current

3.3. Architecture diagram

Open the full-size diagram for a full-screen view.

3.4. Future — Missing pair integrations

No high-confidence opportunities identified.

3.5. Future — Candidate new services

No high-confidence opportunities identified.

3.6. Future — Unused features in this service

No high-confidence opportunities identified.