Comparison — Knovaryn vs Docling SDK / Docling MCP¶
Factual capability comparison, reviewed 2026-08. Capabilities are "documented present / absent as of review"; projects move fast, so verify against the live repos before relying on a claim. See also the peer landscape.
What Docling is¶
Docling (IBM) is a document conversion / parsing library and MCP server for high-fidelity extraction of PDFs and office documents into structured representations (canonical DoclingDocument JSON, Markdown). Knovaryn uses Docling as its canonical parsing layer (ADR 0003), so this comparison is complementary rather than adversarial.
Where they differ¶
| Concern | Knovaryn | Docling (SDK / MCP) |
|---|---|---|
| Scope | Training-data foundry (parse→split→generate→gate→export→publish) | Document parsing / conversion |
| Document → dataset | Yes (end-to-end) | No — parse only (you build the rest) |
| MCP surface | knovaryn_mcp (full dataset workflow: ingest → estimate → run → review → export) |
Docling MCP (parsing tools) |
| Provenance | Span-level enforced lineage + content hash on examples | Returns structured docs for you to process |
| Quality gate | Fail-closed quarantine on generated examples | N/A (parsing) |
| Relation | Consumes Docling's canonical JSON | Provides the parse |
Docling is the right tool when you need the best-in-class doc parsing surface by itself. Knovaryn is the right tool when you want to go all the way from permitted documents to a gated, exported, traceable training dataset — and it leans on Docling for the parsing step.
When to choose Knovaryn¶
- You want a complete document-to-training-data pipeline, not just parsing.
- You want enforced lineage, quality quarantine, and durable jobs on top of the parse.
- You want to drive dataset construction from an MCP-capable agent.
When Docling may fit better¶
- You need document conversion/extraction only, and will build your own training pipeline on top.
No superiority claim is implied — Docling is the parsing dependency that makes Knovaryn's fidelity possible. See the peer landscape for the wider field.