Knowledge graphs and GraphRAG

Search and Retrieval, lesson 13

Store the links between facts, and know when a graph beats vectors.

"Vector search" and "Combining keyword and vector search": finding the passages that look most like a question. "Why grounded answers still go wrong": answers that need two passages at once. This lesson stores the links between facts, so a question can follow them.

Part 1. Answers that sit between passages

An e-bike maker keeps its supplier notes in the same search system as its handbook. Nine of them are below, one sentence each. A buyer hears that Norfield Metals may shut its plants, and asks: "Which products depend on Norfield Metals?"

Example 1.1

Nine notes, and what search returns

n1Norfield Metals supplies our aluminium frames.
n2Brake discs come from Norfield Metals, under a contract renewed in 2025.
n3Norfield Metals has two plants, both in Poland.
n4Kestrel Cells makes the battery packs.
n5Volta Circuits builds the motor controllers.
The notes about suppliers. Each one says who makes which part.
n6The Trail e-bike uses the aluminium frame.
n7The Trail e-bike and the City scooter are both fitted with brake discs.
n8Battery packs go into the Trail e-bike and the City scooter.
n9The motor controller is fitted to the City scooter.
The notes about products, written by another team. None of them names a supplier.
Which products depend on Norfield Metals?
BM25 top 3→n1 2.44n3 2.17n2 1.79
BM25 as in "Keyword search with TF-IDF and BM25" (k1 = 1.5, b = 0.75, punctuation stripped). Only the three notes with "Norfield" and "Metals" score at all, and the short note about Poland beats the one about brake discs.
products named inn6n7n8n9
shared words with the question: 0
No note that names a product shares a word with the question. An embedding search orders things differently, but those notes are about bikes and scooters, not Norfield, so they sit far from the question too.

Your turn

The model gets n1, n3 and n2, plus the question. What's the best answer it can honestly give?

The answer is a chain: Norfield supplies brake discs (n2), and brake discs go into the City scooter (n7). n7 shares no words with the question, so it never arrived. A careful model says what's missing. A careless one guesses "the e-bikes" and leaves out the scooter.

Show the answer

A. The answer is a chain: Norfield supplies brake discs (n2), and brake discs go into the City scooter (n7). n7 shares no words with the question, so it never arrived. A careful model says what's missing. A careless one guesses "the e-bikes" and leaves out the scooter.

"Why grounded answers still go wrong" split a two-topic question into two searches. That needs both topics up front, and here you only learn to search for "brake discs" after reading the first results. A model can search, read and search again, paying extra calls on every question. Or you can record the links once, while indexing, and follow them for almost nothing later.

Multi-hop question

A question whose answer joins facts from two or more passages, where each fact leads to the next. Each step along a link is a hop. "Which products depend on Norfield Metals?" takes two: supplier to part, then part to product.

Chains are one of three kinds of question that top-k search handles badly. The second is about the whole collection: "What are the most common reasons for late deliveries?", asked of 20,000 notes, has an answer spread across all of them. The third is counting. "How many of our suppliers have plants in Poland?" needs every matching note, search hands over at most k, and nothing tells the model how many it missed.

Your turn

The supplier notes grow to 20,000. Which question can top-5 search still answer well?

The plants question has its answer in one note (n3), which shares "Norfield", "Metals" and "plants" with it: one fact, in one passage, in the question's own words. That's what top-k search is built for. The count needs every supplier note and a total no note contains, late deliveries is a whole-collection question, and the products question is the two-hop chain.

Show the answer

B. The plants question has its answer in one note (n3), which shares "Norfield", "Metals" and "plants" with it: one fact, in one passage, in the question's own words. That's what top-k search is built for. The count needs every supplier note and a total no note contains, late deliveries is a whole-collection question, and the products question is the two-hop chain.

Part 2. Facts as triples

Keep going with Pro

You've done the first part. Pro opens the rest of this lesson, and every other one.

Still to come

Facts as triples
Walking the graph
GraphRAG, and when a graph is worth it

See Pro plans I have Pro: sign in

Next lessonCapstone: a search-backed assistantBuild and measure an assistant that answers from your documents. Start lesson 14

Search and Retrieval overviewSpotted a mistake? Tell us.

13 · Knowledge graphs and GraphRAG