Lab component — not in the Synedrix OS tech stack·Evaluated under LAB-002·Not required in any deployment mode

Not a Synedrix OS platform. Docling is not one of the twelve platforms and holds no OS layer. It appears here only as the structural-parsing engine Semantica wraps (semantica[parse-docling], >=2.107.0 per Semantica’s own pyproject.toml) — the one capability LAB-002 actually scoped as “in scope — the main event.” This page is lab reference material under LAB-002, evaluated in structural isolation from production — not a stack platform page, and not citable as authority by any canonical document.

What Docling is

Docling (docling-project/docling, MIT) is a document-conversion library that parses PDFs, Office documents, HTML, and Markdown into a unified structural representation — preserving tables, nested headings, reading order, and layout, rather than flattening a document straight to plain text.

ConceptWhat it is
DocumentConverterDocling’s own top-level entry point — takes a file path or URL, returns a structured DoclingDocument
DoclingDocumentThe structural output: a tree of headings, paragraphs, and tables, each table carrying real row/column structure rather than a flattened text blob
Table Structure Recognition (TSR)The specific capability this lab exists to test — Docling’s layout model detects table boundaries and reconstructs header/row association, confirmed under LAB-002’s stand-up to correctly parse a real 14×7 table
BackendsMultiple parse backends per format (e.g. PDF via its own layout/OCR pipeline); format detection is automatic from file content, not just extension

What Docling is normally used for

  • RAG ingestion pipelines — converting heterogeneous source documents (PDF reports, Office docs, HTML) into a structure a chunker can reason about, rather than raw text a chunker has to re-infer structure from
  • Table-heavy document processing — financial reports, technical specs, anything where a table’s header/row association is load-bearing information that plain-text extraction destroys
  • Multi-format normalization — one converter across PDF/DOCX/PPTX/HTML/Markdown, rather than a different bespoke parser per format

The common thread: Docling’s value is preserving a document’s real structure through the parse step, so whatever consumes its output (a chunker, an embedder, a retrieval pipeline) doesn’t have to reconstruct that structure from degraded plain text.

How Synedrix OS uses it

Synedrix OS does not use Docling in production. It is evaluated only inside LAB-002’s isolated environment, reached exclusively through Semantica’s DoclingParser wrapper class — this lab never depends on Docling directly, and never installs it outside the semantica[parse-docling] extra.

flowchart LR
  n1([Docling]):::term
  subgraph What it owns
    n2[Nothing in production]:::agent
  end
  n1 --> n2
  subgraph What it does not own
    n3[Semantic retrieval and similarity search]:::warn
    n4[Qdrant]:::term
    n3 --> n4
  end
  n1 --> n3
  subgraph Why Docling is evaluated and what would justify adopting the capability
    n5[It is the one concrete testable claim in Semantica s whole capability bundle]:::term
    n6[The mechanism is real and independently verified]:::term
    n7[The result is measured not assumed]:::term
    n8[Adoption if it ever happens is scoped narrowly]:::term
  end
  n1 --> n5

What it owns

Nothing in production. Inside LAB-002’s isolation boundary, DoclingParser runs inside the synedrix-lab-semantica container and structurally parses the same docs/openwiki/academy corpus the production KB ingests, feeding a separate, additive Qdrant collection (synedrix_os_kb_semantica_v1 on qdrant-lab) built specifically to compare against production’s existing chunker.

What it does not own

ConcernOwner
The production chunking/ingestion pipeline itselfThe existing Markdown → Qdrant chunker in synedrix-os-website/agent (production, unmodified by this lab)
Semantic retrieval and similarity search — the single vector substrate for Layer 7Qdrant (Layer 7); Docling’s lab collection is additive, never a replacement
Anything a registered agent may recall as memoryThe governed memory fabric (Memory Fabric Spec), never a lab component directly

Why Docling is evaluated, and what would justify adopting the capability

  • It is the one concrete, testable claim in Semantica’s whole capability bundle. Everything else Semantica ships — RDF triple store, self-implemented graph algorithms, OWL/SHACL ontology — was pre-judged out of scope by the LAB-002 charter as substrate duplication or a bigger question than the lab’s timebox supports. Structural parsing is the exception: a falsifiable, measurable question (does it improve retrieval?), not an architectural judgment call.
  • The mechanism is real and independently verified. DoclingParser correctly extracted a genuine 14×7 table under LAB-002’s stand-up — confirmed by inspection, not assumed from Docling’s own documentation.
  • The result is measured, not assumed. LAB-002’s retrieval-quality harness (recall@k/MRR) found recall@10 on multi-hop queries went 0.25 → 0.75 with Docling’s table-aware chunking, table-heavy and simple-lookup categories staying tied — at n=4 per category, a real first signal, not yet a robust one. See the LAB-002 charter for the full breakdown.
  • Adoption, if it ever happens, is scoped narrowly. Even a positive result would justify adopting Docling’s table-aware parsing as a preprocessing step ahead of the existing chunker — not adopting Semantica, and not any of the capabilities this charter already pre-judged out of scope.

Official documentation

Synedrix OS · Lab component reference · Docling