Building a Graph by Hand

Where we left off

Lesson 4 produced a flat list of triples, ("Mia", "recalibrated", "humidity sensor") and others like it. A list is still not a graph you can traverse, though: to answer "what did Mia do?" from a plain list, you'd have to scan every triple checking whether its subject is "Mia." That's fine for six triples, unbearable for six thousand. This lesson turns the list into an actual graph structure built for fast lookup: an adjacency dict.

What an adjacency dict actually is

The whole idea, stripped to its essence, before any library involvement: a graph is nothing more than a dictionary where each key is a node (an entity), and each value is a list of the edges leading out of it.

graph = {
    "Mia": [("recalibrated", "humidity sensor"), ("pulled", "multimeter")],
    "humidity sensor": [("located in", "greenhouse")],
    ...
}

That's it. No special library, no database, this is exactly what networkx (introduced much later, in the Advanced tier) does underneath, just with more features bolted on. Building it by hand first means that when Lesson 20 finally imports networkx, you'll recognize add_edge as doing something you've already done yourself, not a mysterious black box.

Why store both directions

def add_edge(graph, subject, relation, obj):
    graph.setdefault(subject, []).append((relation, obj))
    graph.setdefault(obj, []).append((f"(reverse) {relation}", subject))

You might assume storing ("Mia", "recalibrated", "humidity sensor") once is enough, since the fact is fully captured. It's captured, but not reachable from both ends. If a later traversal starts at "humidity sensor" (because that's the entity found in the question), and the edge only exists on graph["Mia"], the traversal from "humidity sensor" has no way to discover Mia at all, real relationships go two directions, even when the sentence describing them only reads in one. Storing the reverse edge too, labeled distinctly so it doesn't get confused with a real forward relation, is what makes traversal work starting from either endpoint, which matters a lot once Lesson 6 starts traversing from whatever entity a question happens to mention first.

The code, piece by piece

graph: dict[str, list[tuple[str, str]]] = {}
for subject, relation, obj in triples:
    add_edge(graph, subject, relation, obj)

Nothing surprising: walk the triples from Lesson 4, add each one (and its reverse) into the dict. setdefault(subject, []) is doing the "create this key if it doesn't exist yet" work a defaultdict would also do, spelled out explicitly here since this course hasn't introduced defaultdict yet and the goal is maximum transparency over what's happening.