Conceptual Interchange Format
A language for machine knowledge

CIF is a compact, language-independent representation for conceptual knowledge using typed triples and stable identifiers. Stable IDs, auditable inference, reproducible derivations.

C_DOG TYPE C_MAMMAL (CONF=0.912)
C_DOG PART_OF C_CANIDAE (CONF=0.881)
U_REX INSTANCE C_DOG (CONF=1.000)
# inferred by TYPE_TRANSITIVITY:
C_DOG TYPE C_ANIMAL (CONF=0.832)
22
Operators
117K
Concepts (WordNet)
167K
Places (Geo)
0
Trainable params (GCE)

Early release. The specification is stable; the reference engine and geometric retrieval are reproducible and deterministic — a reference implementation, not a hardened production system. The NL→CIF translation pipeline is under active development. Roadmap ↓

What is CIF?

A language, not a database

CIF defines how to represent concepts, relations, and metadata. 22 typed operators cover identity, hierarchy and mereology, spatial and temporal anchoring, causation, attribution, and comparison. The specification is the single source of truth.

Deterministic inference

TYPE transitivity, PART_OF weakest-link, causal chains, and temporal reasoning fire by default; cross-scope SAME_AS bridges are honoured on demand under the deployment's federation policy. Every inferred triple carries the rule that produced it and a computable confidence score.

Geometric retrieval (GCE)

The Geometric Concept Engine answers subsumption and nearest-region queries from an exact interval tiling computed directly from the type DAG — zero trainable parameters, deterministic. A trained hyperbolic embedding is retained as an alternate backing; the WN18RR link-prediction eval is reproducible from experiments/.

Architecture

User (natural language) | v [cif_translator] IntentStage > StopWordStage > LexiconStage detects intent (QUERY / COMPARE / EXPLAIN / ...) resolves tokens > CIF entity IDs | v [cif_reasoner] rule-based inference TYPE trans. | PART_OF trans. | causal | temporal | v [cif_knowledge] knowledge provider (GCE+JSON | Neo4j | in-memory) | v User (answer + evidence flags [TYPE_TRANSITIVE] [COMPARE] ...)

Quick start

# Clone and install git clone https://github.com/cif-core/cif-core.git cd cif-core python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt # Start the API (WordNet profile — no database required) COMPONENTS_FILE=api/components.wordnet.yaml \ uvicorn api.main:app --host 127.0.0.1 --port 8001 # Interactive docs at http://127.0.0.1:8001/docs
# Or use the CLI directly python scripts/ask.py --kb wordnet "what is a dog?" # With inference python scripts/ask.py --kb wordnet --reason "what is a mammal?"
# Docker docker compose up wordnet # http://localhost:8001 docker compose up geo # http://localhost:8002 docker compose up stars # http://localhost:8003

Knowledge profiles

These profiles are working examples, not limits. CIF is domain-agnostic — you can build your own ontology from scratch, define your own concepts and operators, and load it as a profile.

Switch domains without code changes. Each profile loads a different knowledge base and activates the relevant operators.

Profile Operators Data Size
wordnet TYPE, INSTANCE, PART_OF, MEMBER_OF, HAS_ATTR Princeton WordNet ~117K concepts
geo AT + OFFSET + UNIT (geographic KNN), TYPE, INSTANCE GeoNames ~167K places
stars TYPE, INSTANCE, VALUE Stellar catalog Demo

Who is this for?

Researchers

A formally specified conceptual language with reproducible scripts. Cite the RFC, run the included benchmarks, extend the operator vocabulary through the documented adoption surface.

Developers

A REST API with 25 endpoints. NL queries, closed-world inference, concept lookup, abductive reasoning. Plug in your own knowledge base.

Organizations

Deterministic knowledge infrastructure with no training-data dependency and a full audit trail on every inference. Suitable for settings where reasoning steps must be inspectable and reproducible. Commercial license available.

Roadmap

Component Status What it means
RFC-CIF-0001 v1.0 Final 22 typed operators, inference rules, metadata — stable
Inference engine Reference TYPE, PART_OF, causal, temporal — deterministic; SAME_AS bridges via federation policy
GCE Reference Exact interval tiling — zero trainable parameters, deterministic; trained hyperbolic embedding as alternate backing
REST API Reference 25 endpoints — query, infer, compare, search, reason
NL→CIF translation In development Querying existing KBs works; generating triples from free text is being reworked
Domain profiles Coming Healthcare, legal, industrial — curated knowledge bases
Concept registry Planned Global C-ID assignment — DNS for concepts

The CIF-Core Foundation

CIF is designed to be a public standard, not a product owned by any single entity. The CIF-Core Foundation is being established to maintain the specification, the ID registry, and the reference implementation as public infrastructure.

Neutral

The standard does not belong to any government, corporation, or superpower. Governance follows the W3C model: open process, consensus-driven.

Ethical use

Binding ethical principles are written into the foundation's charter. CIF must serve humanity, not extract from it.

Open specification

The RFC, papers, and reference implementation are permanently open. The foundation guarantees they cannot be locked down.

Get involved

Build a knowledge profile

CIF is domain-agnostic. Define your ontology in YAML, generate a JSON knowledge base, and load it as a profile. Medical, legal, industrial, educational — any domain works.

Extend the engine

Add new inference rules, implement a translator backend, or connect CIF to your existing knowledge graph. The architecture is modular by design.

Join the foundation

We are looking for researchers, developers, and organizations who share the vision of open conceptual infrastructure. Reach out.

Citation

@software{cif_core_2026, title = {CIF-Core: Reference Implementation of CIF (Conceptual Interchange Format)}, author = {Villa, Hector}, year = {2026}, url = {https://github.com/cif-core/cif-core} }