Semantica

Semantica is a Python framework that puts a semantic and context layer underneath an LLM, a vector store, and an agent framework. It ingests enterprise data, extracts entities, relations, events, and triplets, and builds a Context Graph and knowledge graph that carries business meaning rather than embedding similarity alone. Graph construction, reasoning, and provenance are deterministic and need no LLM. Where an LLM is used (for example in extraction or community summaries) it is optional and vendor-neutral.

Every agent decision can be recorded as a first-class graph node with structured context, linked to upstream causes and downstream effects, searched by precedent, and exported as W3C PROV-O. The same structure answers "why did this happen?" for compliance, risk, and audit teams, and the framework explains what happens outside the model (the context fed in, the decision produced, applied policies, and the execution trail), not the model's internal reasoning. It is aimed at regulated domains such as finance, healthcare, legal, government, and defense, and at data teams on Databricks, Snowflake, or SAP who want a governed, lineage-tracked graph without exporting to a third-party service.

The pipeline runs Sources, Ingest, Parse, Normalize, Split, Extract, Conflict Detection, and Deduplication into a knowledge graph, with ontology, reasoning, provenance, and decision modules on top, then stores into vector stores and polyglot graph storage. Storage is swappable without code changes: RDF triple stores (embedded Oxigraph, Blazegraph, Apache Jena, Eclipse RDF4J), labeled property graphs (Neo4j, FalkorDB, Apache AGE, AWS Neptune), and vector stores (FAISS, Qdrant, Weaviate, Milvus, Pinecone, PgVector). Reasoning engines cover forward chaining, Rete, Datalog, and SPARQL with explainable paths, and bi-temporal facts support point-in-time snapshots and Allen interval algebra.

Semantica reaches agents through first-class Agno, CrewAI, and LangChain integrations (LangGraph through the LangChain extra), an MCP server whose tools extract entities and relations, record and query decisions, find precedents, walk causal chains, edit the graph, run reasoning, and export, a REST API, and a CLI with command groups for every module. Plugin bundles with domain skills and specialized agents ship for Claude Code, Cursor, Codex, Windsurf, Cline, Continue, VS Code, OpenClaw, and Pi. LLM access goes through its own provider module and LiteLLM, naming OpenAI, Anthropic, Google Gemini, Mistral, Meta Llama, Groq, Cohere, Azure OpenAI, AWS Bedrock, Ollama, DeepSeek, Perplexity, Together AI, Fireworks AI, Replicate, and Hugging Face. A browser-based Knowledge Explorer pans live graphs, scrubs a timeline, reviews decision chains, merges duplicates, and edits ontologies visually.

Reported benchmarks from v0.5.0 on a 118,000-node graph:

Operation	Before	After	Improvement
Node search (118k nodes)	24 ms	0.004 ms	6,000× faster
Embedding cache hit	cold load	revision-based cache	10× throughput
Semantic deduplication	baseline	optimized candidate generation	6.98× faster
Candidate generation	baseline	blocking strategy	63.6% faster

The deduplication and candidate-generation figures are historical measurements rather than automated test assertions, and results vary by hardware, dataset, and backend.