Extracting Relationships by Hand

Where we left off

Lesson 3 pulled entities out of a document, a list of who and what it mentions. A list of entities alone isn't a graph yet, though, it's just a bag of nouns with no structure between them. What actually makes something a graph is the edges: the relationships connecting one entity to another. This lesson extracts those, as (subject, relation, object) triples, the standard shape a knowledge graph edge takes.

Why triples, specifically

You might assume any free-text description of a relationship would do, "Mia used the multimeter to recalibrate the sensor" is perfectly readable. The reason to force it into (subject, relation, object) instead is mechanical, not stylistic: a triple is traversable. Given the triple ("Mia", "recalibrated", "humidity sensor"), code can ask "what did Mia recalibrate?" or "who recalibrated the humidity sensor?" by pattern-matching on a fixed position, subject or object. A sentence of free text can't be queried that way without re-running language understanding on it every single time you want to ask a different question of it. The triple pays that understanding cost once, at extraction time, so every later traversal is just a lookup.

The code, piece by piece

prompt = f"""...list every relationship between two entities as
a (subject, relation, object) triple...

Return ONLY a JSON array of 3-element arrays, like:
[["Mia", "recalibrated", "humidity sensor"], ...]

Text:
{text}"""

Same JSON-mode pattern as Lesson 3, applied to a slightly harder extraction task: instead of "list the nouns," this asks the model to also identify the verb phrase connecting two of those nouns, and to keep that verb phrase short and consistent (a few words, not a full sentence), so triples about the same kind of relationship end up comparable to each other later.

triples = extract_relationships(text)

Notice what this does not do yet: it doesn't merge these triples with any other document's triples, or build an actual graph structure out of them. That's Lesson 5. This lesson's whole job is turning one document's prose into a flat list of (subject, relation, object) triples, nothing more.