Combining Graph and Vector Retrieval

Where we left off

Every Beginner-tier lesson picked its starting entity by hand: START_ENTITY = "humidity sensor", typed directly into the script, because the fixtures are small enough that a human can just read them and know where to start. A real system doesn't get that shortcut, it receives a question as plain text and has to figure out, on its own, which graph node to start traversing from. This lesson fixes that, using a tool this course's readers already know well: embeddings.

Why embeddings are the right tool for this, specifically

You might assume finding a starting node just means checking whether any graph node's name appears as a literal substring of the question. That works embarrassingly rarely: this lesson's own example question below says "the sensor," not "humidity sensor," a phrasing gap a substring check can't bridge but an embedding comparison can, since "the sensor" and "humidity sensor" land close together in meaning even though they don't share the same characters. This is precisely naive_rag Lesson 3's cosine-similarity idea, reused for an entirely different job: instead of ranking chunks by similarity to a question, this ranks graph nodes by similarity to a question, then hands the single best match to Lesson 7's traversal as its starting point.

The code, piece by piece

node_names = list(graph.keys())
node_vectors = embed_texts(node_names)
query_vector = embed_texts([query])[0]

best_node = max(
    node_names,
    key=lambda name: cosine_similarity(query_vector, node_vectors[node_names.index(name)]),
)

Embed every node name once (a single batched call, the same batching habit naive_rag Lesson 5 established), embed the question once, then pick whichever node's embedding is closest to the question's. This best_node becomes the traversal's starting point, no hand-typed START_ENTITY constant needed anymore.

facts = gather_facts(graph, best_node, max_hops=3)
answer = generate_answer(query, facts)

Past choosing the starting node, this is Lesson 7's traversal and generation, unmodified, with one adjustment worth noticing: max_hops goes up to 3 here, from Lesson 7's 2. A vector-picked node is a semantic match to the question, not necessarily the single closest node to the answer in the graph itself, running this lesson's own example shows the closest embedding match is "greenhouse", a real and reasonable match, but one hop further from Mia and the multimeter than Lesson 7's hand-picked "humidity sensor" was. Automating the starting point trades a little precision in where you start for not needing a human to pick it, and this lesson pays that cost back with one extra hop. Lesson 14 makes that tradeoff, and its downside, explicit.