04 · What an ontology gives an AI agent

Ontology video

Why your AI agent can’t connect the dots. “Supplier Acme is late. Which customers are affected?” The answer is spread over three systems: purchasing, product data and orders. An ontology names the links between them, so the agent can follow the chain: supplier → part → product → customer. Three hops later: Kestrel and Orbis, with the path as proof.

Watch: YouTube · TikTok · Instagram (92 seconds, plus a 34-second short)

The idea

Facts are triples: (subject, relation, object), like ("Acme", "supplies", "bolt").
The ontology says what kinds of things exist and how they may connect: a Supplier supplies a Part, a Part is used_in a Product, a Product is ordered_by a Customer.

With the ontology, a multi-hop question becomes a walk along named relations. Without it, an agent searching separate tables for “Acme” finds the supplier row and stops there.

Run it

python3 src/ontology.py

supplies -> ['bolt', 'hinge']
used_in -> ['Drone', 'Locker']
ordered_by -> ['Kestrel', 'Orbis']
affected: ['Kestrel', 'Orbis']

The short version (short/ontology.py, 7 lines) prints only the last line.

Files

04-ontology/
├── data/
│   └── facts.csv         10 facts: subject, relation, object
├── src/
│   ├── ontology.py       the code from the video
│   └── facts.py          loads facts.csv
└── short/
    ├── ontology.py       the 7-line version from the short
    └── facts.py

The long and short versions read the same data/facts.csv.

facts.csv holds 10 triples, the kind of rows you’d pull from an ERP (purchasing), a PLM (product data) and a CRM (orders).

The code

src/ontology.py, line by line:

Line	Code	What it does
1	`from facts import FACTS`	The instance data: 10 triples.
3 to 6	`ONTOLOGY = [...]`	The schema: three typed relations that chain Supplier → Part → Product → Customer.
8	`def follow(start, path):`	Walk from a starting entity along a list of relations.
9	`found = {start}`	Begin at the late supplier.
10	`for rel in path:`	One hop per relation.
11 to 12	`found = {o for s, r, o in FACTS if s in found and r == rel}`	Everything reachable from what we’ve found, using this relation.
13	`print(rel, "->", sorted(found))`	Show each hop. This is the “path as proof”.
16	`path = [rel for _, rel, _ in ONTOLOGY]`	The relation chain, read straight from the ontology.
17	`print("affected:", follow("Acme", path))`	Ask the question.

Vela is not affected: it orders the Pump, which uses a gasket from Zenco, not Acme.

Why it matters

Real companies run on many systems that name the same things differently. An ontology gives an agent shared meaning across them. The agent hallucinates less, because it follows real links instead of guessing, and every answer can show its path, which makes it auditable.

Try this

Make Zenco the late supplier. Who is affected?
Add a fact: ("bolt", "used_in", "Pump"). Does Vela become affected?
Add a fourth relation, like ("Customer", "located_in", "Region"), and ask which regions are affected.
Change follow to also return the full path for each customer, not just the final set.

Previous: 03 · The AI agent loop · Next: 05 · Evals and loop engineering · All lessons