Semantica

A Python library that builds a knowledge graph out of enterprise data and keeps a record of what an agent decided and why. MIT, pip install semantica, from semantica-agi (founder and main committer mohd-kaif). At the read on 2026-08-18: 8,719 stars, 894 forks, 2,357 commits, 101 open issues, first commit 2025-06-25, last push that same morning.

The README calls it "Graph-Native Infrastructure for Context and Accountable AI Systems" and, one line down, "The Open Source Palantir for AI Agents." The target buyer is a regulated one — lending, healthcare, legal, government — and the pitch is aimed at a question those buyers get asked: "an underwriting agent's approval has to survive a regulator's 'why' months later."

Where it sits

Underneath the LLM, not beside it. "Semantica sits underneath your LLM, vector store, and agent framework as a deterministic infrastructure layer: no LLM required for graph construction, reasoning, or provenance." The pipeline is ingest → parse → normalize → split → extract → conflict detection → deduplication → knowledge-graph → ontology / reasoning / provenance / decisions → RDF and property-graph storage → export, REST, MCP, CLI. Every stage ships as an importable module.

Storage is deliberately plural: RDF triple stores (Oxigraph, Blazegraph, Jena, RDF4J) and labeled property graphs (Neo4j, FalkorDB, Apache AGE, Neptune), plus the usual vector stores, swappable without code changes. Ingestion reaches Databricks Unity Catalog and Snowflake directly, which is the part of the design that reads as enterprise-first rather than research-first.

Decisions as nodes

The distinguishing feature is that a decision is a graph object, not a log line. record_decision() stores category, scenario, reasoning, outcome and a confidence number; add_causal_relationship() wires it to upstream causes and downstream effects; trace_decision_chain() walks the ancestry back; find_similar_decisions() searches precedent. Audit trails export as W3C PROV-O, chosen because that is the format compliance frameworks accept.

Reasoning is deliberately old-fashioned — forward chaining, a Rete network, Datalog, SPARQL — so the explanation is a derivation rather than a generated paragraph. The README is careful about the limit of that claim, and unusually so for a project README: "System-level explainability, not foundation-model explainability." What happens inside the model stays opaque; what Semantica explains is the context going in, the decision coming out, and the trail between them.

The temporal design is bi-temporal — valid time separate from recorded time — with point-in-time snapshots, the same shape graphiti uses for agent memory and temporal-knowledge-graph describes. Conflicting facts are flagged rather than overwritten, which is this corpus's own record-don't-overwrite rule implemented in a database.

The numbers it publishes are speeds

From v0.5.0, on a 118,000-node production graph (AMD EPYC, 64 GB): node search 24 ms → 0.004 ms, semantic deduplication 6.98× faster, candidate generation 63.6% faster via a blocking strategy. The footnote is honest about their status — historical CHANGELOG measurements, not asserted by a test — and points you at pytest tests/vector_store/test_performance_benchmarks.py to measure your own data.

Every one of those figures is a speed. Nothing in the README says how many merges were right. See entity-resolution for why that matters here: this is the largest deployed graph in the corpus that ships blocking plus semantic dedup, and it still leaves the corpus's open question untouched.

Reading it as a source

Everything above is self-reported. There is no third-party benchmark, no evaluation against graphrag-family systems, and the comparison table that scores "Vector DB + RAG" and "Plain LLM Memory" as None on provenance, reasoning, conflict detection and time travel is a vendor's table, not a measurement — graphrag-bench exists precisely because that genre of claim needs a neutral referee. Tiered T1 as first-party documentation of what the software does; its comparative claims are advocacy.

Cross-spoke: the RDF/OWL/SHACL/SPARQL/SKOS machinery belongs to knowledge-representation's spoke (knowledge-representation-wiki), and the compliance-and-audit framing touches ai-governance-wiki. Routed here because the substance is graph construction for AI context and memory — the graphiti / lightrag / graphrag cluster.

Related

knowledge-graph · entity-resolution · temporal-knowledge-graph · graphiti · graphrag · graphrag-bench · agent-memory-knowledge-graphs · graph-construction-quality · ontology · semantica-agi · mohd-kaif