Semantica — Lab Component

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

Not a Synedrix OS platform. Semantica is not one of the twelve platforms and holds no OS layer. It appears nowhere in the canonical document set prior to LAB-002 — unlike EverOS, there is no prior ADR position to respect or discharge here; this is a first look at a set of independent capabilities, not a follow-through. This page is lab reference material, evaluated in structural isolation from production — not a stack platform page, and not citable as authority by any canonical document.

What Semantica is link

Semantica (semantica-agi/semantica) is a Python knowledge-graph and GraphRAG toolkit — real, active, MIT licensed, at PyPI version v0.6.5. It bundles several genuinely independent capabilities rather than one narrow function.

What Semantica is normally used for link

• GraphRAG retrieval — building a knowledge graph over a document corpus so multi-hop questions (tracing every document a decision touched, following a causal chain) become answerable, which pure vector similarity search structurally cannot do
• Decision and provenance tracking — recording a decision, its reasoning, and its outcome as a queryable graph node, then exporting that lineage in a standards-compliant format
• Structural document ingestion — extracting tables and layout faithfully from source documents before chunking, rather than treating a document as an undifferentiated block of text

How Synedrix OS uses it link

Synedrix OS does not use Semantica in production. LAB-002 evaluates a deliberately narrow subset of its capabilities, scoped by what is genuinely undecided rather than testing everything it bundles equally.

flowchart LR
  n1([Semantica]):::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
    n5[Canonical relational truth]:::warn
    n6[Supabase PG17]:::term
    n5 --> n6
    n7[Explicit relationships and lineage as production authority]:::warn
    n8[Neo4j]:::term
    n7 --> n8
  end
  n1 --> n3
  subgraph Why Semantica is not yet in the stack
    n9[Pre-1.0 despite a Production/Stable classifier]:::term
    n10[The RDF/graph-algorithm capabilities duplicate what the OS already has]:::term
    n11[The one genuinely novel claim is not yet verified]:::term
    n12[Worth evaluating precisely because nothing here is decided yet]:::term
    n13[Docling parsing works with one real tempering finding]:::term
    n14[PROV-O export is real and genuinely standards-compliant once the correct API]:::term
    n15[The Neo4j graph bridge is real but was initially unused]:::term
  end
  n1 --> n9

What it owns link

Nothing in production. Inside LAB-002’s isolation boundary, a Semantica-built Qdrant collection and a Semantica-built decision graph in neo4j-lab are both reachable only from lab infrastructure, never from a registered agent, and hold no production tenant data.

What it does not own link

Why Semantica is not (yet) in the stack link

• Pre-1.0 despite a Production/Stable classifier. Commit history is heavily concentrated — 1,823 of the top contributor’s commits versus 239 for the next-largest, across 54 nominal contributors — a materially less mature project than its PyPI classifier implies.
• The RDF/graph-algorithm capabilities duplicate what the OS already has. The same substrate-duplication rule that rejected EverOS’s runtime applies directly: a triple store duplicates Neo4j, and Semantica’s graph algorithms are self-implemented Python (NetworkX/Gensim), not a Neo4j GDS integration — strengthening, not weakening, the case for leaving both out of scope.
• The one genuinely novel claim is not yet verified. Whether Docling’s structural parsing measurably improves retrieval over the existing chunker requires a recall@k/MRR comparison on a labelled query set, split by query type — not yet built. A populated comparison collection existing is not the same as the comparison having run.
• Worth evaluating precisely because nothing here is decided yet. Unlike EverOS, no prior ADR has ruled on Semantica — LAB-002 is a first look, and a negative result on any capability above is exactly as valid an outcome as a positive one.

Real findings from standing it up (Phase 8 onward, August 2026) link

Not assumed from documentation — Semantica was actually built, run, and exercised against real repository content:

• Docling parsing works, with one real tempering finding. A 14×7 table in this repository’s own Tech Stack document was extracted perfectly, header row intact. A different, clean, well-formed 6-row table in the same document was parsed as three separate fragments by Docling itself — Semantica’s own dependency is not immune to the table-fragmentation problem this lab exists to test.
• PROV-O export is real and genuinely standards-compliant, once the correct API (ProvenanceManager, not ContextGraph.export_rdf()) is used — confirmed output includes proper prov:Entity, prov:Activity, prov:Agent, prov:wasGeneratedBy, and prov:qualifiedGeneration triples.
• The Neo4j graph bridge is real but was initially unused. ContextGraph never touches a database on its own; semantica.graph_store.neo4j_store.Neo4jStore is a genuine Bolt driver, subsequently exercised end-to-end by pushing a real four-decision graph (drawn from this repository’s own change logs) into neo4j-lab — 20 nodes, 21 relationships, correctly connected.
• Generic NER quality is genuinely mixed on this corpus. With no spaCy model installed, extraction falls back to naive pattern-matching and misses everything meaningful. With the spaCy English model installed, it catches company/product-shaped entities (Qdrant, Supabase, LangGraph) but misses or mangles this corpus’s own domain identifiers — LanceDB, SQLite, and n8n were missed entirely; ADR-OS-002 was truncated to bare ADR. A reliable knowledge graph over this specific corpus needs pattern-based extraction for its own identifier conventions (ADR-OS-\d+, LAB-\d+, os-agt-\d+), not generic NER alone.
• The embedding provider cannot be pointed at Ollama. semantica.embeddings.OpenAIStore reads only an api_key, with no base_url override — confirmed by reading its source. The ingestion A/B instead reuses the same OllamaEmbeddings/nomic-embed-text call the production pipeline already uses, bypassing Semantica’s own embedding module entirely.
• The MCP server is stdio-only — the original lab compose file’s port mapping for it was wrong and has since been corrected; it now runs idle for docker exec-based use.

Full findings, including the pre-registered exit-evidence questions this evaluation is scored against, are in LAB-002.

Deployment and licensing link

Self-hosting. Semantica is a pip package with optional extras ([parse-docling], [graph-neo4j], [explorer]) — no first-party Docker image exists; the lab builds its own container per-package. The MCP server, when used, is subprocess-spawned by a client rather than run as a standalone service.

Licence — verified. Semantica is licensed under the MIT License, confirmed directly from the LICENSE file in the official semantica-agi/semantica repository.

Official documentation link

• semantica-agi/semantica repository
• Semantica on PyPI
• Semantica LICENSE — MIT License

Related Synedrix OS documents link

• LAB-002 — Semantica Lab Charter — Full evidence, capability-by-capability scope, pre-registered exit criteria, and Phase 8 findings.
• ADR-OS-002 — Memory Fabric Control Plane — §4.2/§4.4, the existing provenance fields Semantica’s PROV-O export is compared against.
• ADR-OS-004 — Lab Lane — The governance mechanism this evaluation runs under.
• Neo4j — The production graph store Semantica’s RDF/graph-algorithm capabilities would duplicate.
• Qdrant — The single vector substrate; Semantica’s ingestion A/B is read-only against it and writes only to its own additive lab collection.

Synedrix OS · Platform reference · Lab · Semantica