cairn-engine

The deterministic entity / world-model engine for AI agents. (The brand is Cairn; the distribution is cairn-engine and the import package is cairn_engine.)

Resolve messy human references to stable canonical IDs, traverse their relations bi-temporally, and author the frozen alias table that makes it all reproducible — storage-agnostic, local-first, zero generative-LLM and zero network on the default path, zero runtime dependencies (pure stdlib).

cairn-engine is the foundational entity subsystem. The agent-retrieval layer (gate, signals, fusion, assembler) lives in the separate cairn-retrieval package, which depends on this one.

Status: feature-complete, pre-release (0.1.0) — full resolution cascade, bounded bi-temporal traversal, ontology authoring + 3-tier dedup, SQLite and openCypher adapters, and a cross-system entity index. 120 tests, ruff + mypy strict, byte-stable. The retrieval-quality benchmark (OP-31) is the next phase; until then, treat determinism/portability as proven and quality as unclaimed. See the roadmap.

Install

pip install cairn-engine

Quickstart

from cairn_engine import (
    Entity, InMemoryAliasTable, InMemoryGraph, Ref, Relation, resolve, traverse,
)

entities = [
    Entity("account::schwab_main", "Schwab", "config", aliases=("my Schwab account",),
           refs=(Ref(doc_id="doc::accounts"),)),
    Entity("doc::payment_retry_policy", "Payment Retry Policy", "document",
           refs=(Ref(doc_id="doc::retry"),),
           relations=(Relation("references", "account::schwab_main"),)),
]

table = InMemoryAliasTable.from_entities(entities)     # build -> freeze -> read-only
graph = InMemoryGraph.from_entities(table.canonical_entities())

resolved, unresolved = resolve("does the payment retry policy touch my Schwab account?",
                               table=table)
# both entities found by gazetteer scan; unresolved == []  (a miss is never a made-up ID)

result = traverse(resolved[0].canonical_id, graph=graph, depth=2)
# connected DocumentRefs, hop-scored 1/(1+hop), byte-stable

The resolution cascade: exact → normalized (NFKC) → fuzzy (token-sort ≥ 0.85, entropy-gated) → semantic (opt-in: your embedder) → LLM arbiter (opt-in: your callable; picks among presented candidates in the 0.70–0.85 band — never mints an ID). Misses come back in unresolved — never a synthesized ID (closed world).

Build-time authoring (offline, human-gated): author_from_text → dedup_candidates (exact → MinHash/LSH → review band, never auto-merged) → your approval → freeze(). Optional extractors (GLiNER / LLM) enter as callables you supply — the engine never imports third-party packages or calls APIs.

Also here: bi-temporal state (as_at, supersede), a cross-system entity index (CrossSystemIndex), a SQLite alias-table backend, and an openCypher compiler (compile_traversal / traverse_cypher) that runs the same bounded traversal on Neo4j / FalkorDB — the engine imports no DB driver; you pass the connection.

Design foundations

Grounded in five CS theories (canonicalization over equivalence classes, labeled property graphs, the bi-temporal model, bounded transitive closure, and the closed-world assumption) — see docs/patterns/ and docs/research/foundations/. Every module is developed intent-first (IVD): constraints in intents/ map 1:1 to tests with hand-computed golden fixtures.

License

MIT — Copyright (c) 2026 Leo Celis