End-to-End Graph QA

Where we left off

Lessons 3 through 7 each built one piece: extract entities, extract relationships, build a graph, traverse it, generate an answer. Lesson 7 ran that pipeline against three hand-picked fixture files. This lesson removes the hand-picking: it extracts from all six of this course's fixture notes into one graph, then answers the multi-hop question by traversing that full graph, the same way a real system would, where you don't get to curate which documents are "the relevant ones" ahead of time, extraction has to happen over everything.

What's actually new here

Nothing conceptually. This is deliberately a consolidation lesson, not a new-idea lesson, the same role naive_rag Lesson 8 played for that course. The one thing worth noticing: extracting from all six documents means the graph now also contains Priya's book club, the workshop's lumber log, and the soil moisture project, entities entirely unrelated to the humidity sensor question. Traversal from "humidity sensor" still only follows the edges that are actually there, so none of that unrelated content shows up in the gathered facts unless a real edge connects to it. That's the graph doing its job: growing the corpus doesn't dilute a specific traversal's precision the way it would dilute a naive top-k similarity search across a larger, noisier document set.

The code, piece by piece

def build_graph_from_notes(notes_dir: Path) -> Graph:
    graph: Graph = {}
    for path in sorted(notes_dir.glob("*.md")):
        text = path.read_text()
        for subject, relation, obj in extract_relationships(text):
            add_edge(graph, subject, relation, obj)
    return graph

This one function is Lessons 3-5 collapsed together: read every fixture, extract triples from each, add them all into one shared graph. Extraction runs once per file, six calls total, and every file's triples land in the same dictionary, so an edge extracted from maintenance-log.md and an edge extracted from electronics-bench.md that happen to share a node (like "multimeter") are now connected in one graph, even though neither file, alone, mentions the other file's content.