Hermes · Adoption Review · issue #86

Docling: adopt as the default, escalate on the hard docs.

T-deep, 5-lane review of Docling as the OCR + document-ingestion engine for #86 (Viska client n8n + pantheon-ops internal). Every claim primary-sourced. The verdict reproduces the production baseline already in our knowledge — The AI Automators run Docling → local VLM → Chandra 2.

Method research-orchestrate · 5 Sonnet lanes Sources primary-verified, dated Repo research/docling-adoption-review-2026-07-01/ Date 2026-07-01

TL;DR

Six bullets

License clean — core MIT, granite-docling-258M weights Apache-2.0 (redistribution-safe). IBM-built → LF AI & Data incubation, 62.5k★, ~daily releases, Production/Stable.
Fits the footprint — CPU-viable, fully offline, ~2.4 GB RAM, 0.6–2.45 pages/sec. No GPU for the standard pipeline. Resolves the #86 self-host concern.
n8n wiring solved — self-host docling-serve-cpu Docker (:5001) ← n8n HTTP Request → POST /v1/convert/source. Skip the v0.0.4 community node for Viska prod.
Reproduces the proven baseline — Docling → local VLM → Chandra 2. We adopt a known-good stack, not an invented one.
The catch — weak on merged/page-span tables, handwriting, scanned, LaTeX, multi-column. Docling's own OmniDocBench composite is UNKNOWN; MinerU/paid APIs beat it there.
Cost — effectively free self-host below ~1M pages/mo; escalation-tier APIs (Mistral OCR batch $1–2/1k) used only for the fraction Docling fails.

The four green lights

◆ Identity & License PASS

Component	License	Notes
docling core · docling-parse · ibm-models (code)	MIT	IBM copyright
granite-docling-258M weights (current VLM)	Apache-2.0	redistribution-safe
SmolDocling weights (superseded)	CDLA-Permissive-2.0	avoid — use granite-docling

IBM Research Zurich → LF AI & Data Incubation (Apr 2025). 62.5k★, 876 contributors, ~1–3 day release cadence, v2.108.0. watsonx managed SaaS ($4/1k pp) signals commitment; OSS core stays MIT.

◆ Capabilities · Architecture · Footprint

Axis	What
Inputs	20+: PDF (digital+scanned), DOCX/PPTX/XLSX, HTML, MD, images, audio/video (ASR), email, XBRL/JATS/USPTO XML
Outputs	Markdown · lossless JSON (DoclingDocument) · DocTags · HTML
Engine	layout (RT-DETR) · TableFormer · reading-order · formula/code · 9 pluggable OCR engines · Granite-Docling VLM · HybridChunker for RAG
Perf	x86 CPU ~3.1s/pg median · M3 Max ~0.32s · L4 GPU ~0.11s · RAM ~2.4 GB · offline, no cloud dep

The catch — where Docling loses

Escalation-tier triggers (route these away from Docling)

Failure class	Evidence	Route to
Merged / page-span tables	issues #2756, #2790 (hallucinated columns)	Chandra 2 · Azure DI · MinerU
Handwriting	#2395 near-total fail; EasyOCR ~64%	Chandra 2 · Azure DI · Textract
Scanned (high-accuracy)	OCR layer < dedicated APIs	Mistral OCR 4 · Azure DI
LaTeX formulas (complex)	<70% BLEU	MinerU (>90%) · Mistral OCR
Multi-column · compliance	#2067 · no SOC2/HIPAA certs	LlamaParse · Azure DI

Benchmark honesty: OmniDocBench — MinerU2.5-Pro 95.75, Mistral OCR 85.66, Marker 78.44; Docling composite = UNKNOWN (unretrievable — flagged, not guessed; opendataloader-bench alt 0.877). MinerU is stronger but AGPL-3.0 → rejected as an engine in the client product. And OmniDocBench is saturated for clean PDFs → the real-corpus benchmark is the gate, not a leaderboard.

Cost · per 1,000 pages

Engine	Cost/1k pp	Accuracy on hard docs	Privacy	Note
Docling (self-host)	~$0–1	Moderate — fails hard classes	Full local	the default
Chandra 2 (self-host)	~$0–1	High — multilingual/forms/HW	Full local	primary escalation, Apache-2.0
Mistral OCR 4 (batch)	$2	Good (OmniDoc 85.7)	API	zero-ops fallback
Azure DI (Layout)	$10	High + compliance	API (BAA)	compliance path
MinerU 2.5	~$0–1	Best OSS (95.75)	Full local	AGPL — internal only
LlamaParse Agentic+	$56	High (GPT-4V/page)	API	expensive

Adopt-path

PHASE 0

Stand up

docling-serve-cpu Docker + n8n HTTP → /v1/convert/source. Smoke on 10–20 real Viska docs.

~½ day

PHASE 1

RAG wiring

HybridChunker (tokenizer = our embedding model) → vector store. Keep md+json lossless.

~1 day

PHASE 2

Escalation tier

Route failure classes → Chandra 2 (self-host) primary, Mistral OCR batch fallback.

1–2 days

PHASE 3 · GATE

Real-corpus benchmark

Docling vs specialist on labeled Viska slice. Scale-up rests on this, not leaderboards.

gate

Open items to route: (a) confirm Viska corpus composition — born-digital vs scanned %, table/handwriting density — drives how heavy the escalation tier must be; (b) GPU on the ops box (only needed for VlmPipeline or self-host Chandra 2 at volume). Feeds the pending viska-n8n + pantheon-ops per-use-case evals.