Graph model selector

Graph Model Selector

Overview

AI agents are only as intelligent as the knowledge structures you give them, and the model you choose constrains what they can reason about. The chapter is explicit: start with your reasoning requirements, not your data. Formal logical inference (medical diagnosis) demands a different model than traversal-heavy network analysis or many-entity events.

Three model classes, each with a characteristic profile:

• Labeled property graph (LPG): fast traversals, mature ecosystem (Neo4j / Neptune / ArangoDB), flexible. Can only tell you what you explicitly programmed — no native inference.
• RDF (subject-predicate-object triples): formal logical semantics enable native inference. Tell it Disease1 causes Symptom1 and Patient exhibits Symptom1 and it infers Disease1 is a candidate diagnosis with no rule written. Slower; n-ary needs reification.
• Hypergraph: one hyperedge connects any number of entities, so n-ary relations are native — no auxiliary nodes. Ecosystem is immature.

The selector scores each model across the five features the chapter names (formal reasoning, n-ary, performance, tool ecosystem, constraint expressiveness) weighted by the caller's requirements, and surfaces a hybrid recommendation when the top two are close (the chapter's "Putting it all together" hybrid guidance).

The n-ary helper makes the trade-off concrete. A prescription (doctor + patient + medication + dosage + date + condition) is one hyperedge (representation_cost.hypergraph_elements == 1) but costs 1 + arity elements (one relation node + one edge per participant) under LPG or RDF reification.

When to Use

• Choosing a graph backend model class for a new agentic system
• Justifying a build-vs-buy or LPG-vs-RDF-vs-hypergraph decision in a design doc
• Deciding whether a multi-entity fact warrants a hyperedge or reification

Phrases: "which graph model", "property graph vs RDF", "hypergraph", "n-ary relation", "reification", "graph data model selection", "Neo4j vs RDF".

When NOT to Use

• Picking a specific product. This selects the model class, not Neo4j-vs-TigerGraph-vs-Neptune. Vendor choice is downstream (see Ch3 build-vs-buy six-factor list).
• Non-graph storage. If the data is genuinely tabular with no relationship reasoning, a relational store is correct.
• An org-mandated model. If finance mandates RDF/OWL or the platform is fixed, adopt it; the scoring is moot.
• A single binary relationship. Arity-2 facts don't need the n-ary helper; a plain labeled edge is fine.

Process

Step	Input	Action	Output	Verification
1	Reasoning requirements (5 weights 0..3)	lib.score_models(reqs)	[(model, score), ...] sorted desc	weights * feature-scores, descending order
2	Same	lib.recommend_model(reqs)	{recommended, scores, rationale, hybrid_recommended}	RDF wins on formal, hypergraph on n-ary, LPG on perf+tooling
3	An n-ary fact as a HyperEdge(type, nodes, attributes)	lib.representation_cost(edge)	{arity, hypergraph_elements=1, property_graph_elements=1+N, rdf_reified_triples=1+N}	hypergraph cost is 1; others are 1+arity
4	Same hyperedge	lib.reify_as_property_graph(edge)	1 intermediate node + N edges listing	intermediate_nodes==1, auxiliary_edges==arity

Rationalizations

Agent rationalization	Documented rebuttal
"Property graphs are popular, just use Neo4j for everything."	Popularity is the tool-ecosystem axis only. If the agent needs native inference (medical diagnosis, compliance), LPG forces you to hand-code every inference rule. The chapter's whole point: the model constrains reasoning — pick from requirements, not from market share.
"RDF does reasoning, so always pick RDF."	RDF pays in performance and n-ary complexity (reification). Traversal-heavy real-time analytics on RDF can be the wrong call. The five-feature scoring exists precisely so one axis doesn't dominate by default.
"I'll force this 6-entity prescription into binary edges — hyperedges are exotic."	representation_cost shows the bill: 1 relation node + 6 edges, and queries must traverse all of them to reconstruct the event. The hyperedge keeps semantic unity in one element. The chapter calls binary-forcing "artificial complexity."
"The selector said hybrid — that's a cop-out."	The chapter explicitly recommends hybrids ("a property graph for high-performance traversals combined with an RDF/OWL layer for formal reasoning"). When two axes both matter and no single model covers both, the hybrid IS the answer — record it as a conscious trade-off, not a non-decision.
"Skip the model question, use PostgreSQL triples (Example 3-2)."	Valid for retrieval-focused systems with predictable query patterns — the chapter says so. But for temporal reasoning, contradiction management, and incremental processing, dedicated graph models win. Score the requirements first; the PG-triples fallback is a deliberate choice, not a default.

Red Flags

• All five requirement weights set to 3. You have not actually prioritized — re-interview the use case. If everything is critical, nothing is, and the selector degenerates to the raw feature averages.
• Recommended = RDF but performance weight is the highest. Mismatch: re-check the weights. RDF's performance score is the lowest of the three.
• Hybrid flagged but the team has no appetite for two systems. Record the trade-off explicitly: pick the single model that covers the higher-weighted axis and accept the documented gap on the other.
• Choosing hypergraph for a team with no hypergraph DB experience. The ecosystem is immature; budget for custom tooling or reconsider.

Non-Negotiable Verification

1. Run the benchmark battery. python cli.py benchmark must report 10/10:
   • RDF wins formal-only, hypergraph wins n-ary-only, LPG wins perf+tooling
   • n-ary cost: hypergraph = 1 element, LPG/RDF = 1 + arity
   • reification produces exactly 1 intermediate node + arity edges
   • hybrid surfaces on balanced formal+n-ary requirements
2. Run the scenario. python cli.py scenario medical-diagnosis recommends an inference-capable model and prints the n-ary prescription cost.
3. Verify CLI help. python cli.py --help exits 0 and prints this SKILL.md description (so any harness can discover the skill from --help).

Security Posture

• Prompt injection. Inputs are requirement weights and hyperedge specs - pure data scored against fixed feature tables. Adversarial values (inflated weights, crafted attribute strings) can at most skew a recommendation; nothing in the input is executed or interpolated into a query.
• Data exfiltration. No network calls, no file writes. Hyperedge attributes may describe sensitive domain facts (the prescription example); they stay in-process and appear only in the stdout cost report the caller owns.
• Privilege escalation. No shell invocation, no eval, no dynamic import. The recommendation is advisory - it selects a model class on paper; provisioning an actual graph backend happens elsewhere under the platform's own controls.

Source Attribution

Distilled from Agentic GraphRAG (O'Reilly, by Anthony Alcaraz and Sam Julien) Ch3 — Knowledge Representation, sections "Graph Data Models", "Evaluating Graph Models" (the five implementation features), and the n-ary discussion with Example 3-1 (hypergraph prescription). The four-stage graph-RAG framing is Zhou et al., "In-depth Analysis of Graph-based RAG in a Unified Framework", cited in the same chapter. PostgreSQL-triples fallback: Example 3-2.