Building a retrieval layer that follows connections instead of matching resemblance

Introduction

Your retrieval works and your answers are still wrong. You ask “what breaks if I change this”, “why did this start failing”, “what are we exposed to” — and get back documents that resemble the question instead of the facts connected to it.

Why tuning the retriever does not fix it

The reflex is to improve the search: a better embedding model, a reranker, a higher k, hybrid keyword-plus-vector. Those help when the answer is ranked low. They cannot help when the answer is ranked at zero — and on this kind of question it is, because the document holding the answer was written by someone who was not thinking about your question and shares no words with it.

In the three labs in this volume, the document that states the answer scores exactly zero against the question every time. No threshold promotes it. That is not a tuning problem, and recognizing it as a different problem is most of what this volume is for.

What graph engineering is

Graph engineering is the practice of converting documents into an explicit structure of entities and relations, so that retrieval can follow connections instead of matching resemblance.

Instead of storing documents and returning documents, you extract the facts inside them as triples — one subject, one relation, one object — and store how those facts connect. A question then starts at an entity and walks, one relation at a time, to facts no document mentioned alongside it.

Where it sits in what you already run

It does not replace your retriever. Finding where to start is still a similarity problem, and this volume keeps it that way — a graph layer takes over after the first step, not instead of it.

So the honest framing is not “graphs instead of vector search”. It is: your search picks the starting point, and the graph decides what travels with it.

Is this for you?

It depends on the shape of your questions, not on the size of your corpus. Sort a few of your real ones:

What it costs

Said before you are invested rather than after. A graph costs a vocabulary somebody maintains, an extraction pass over every document, and a rule for deciding when two names mean one thing. That last one is the hard part, and step 5 is about the two opposite ways it fails.

What you will build

A graph retrieval layer in eight stages, of which five are decisions you make and three are infrastructure you buy. Three runnable labs carry it — a test-failure triage, a code-and-git question, and a supply-chain exposure question. They share their core byte for byte, which is how the claim that the method is domain-independent is made checkable rather than asserted.

By the end you can tell which of your own questions need connection rather than resemblance, build the pipeline that answers them, and say what it cost.

What is original here, and what is not

The retrieval half of this volume is its own: the budget sweep, the hop sweep, the baseline comparison, and every measured figure. The construction half — schema design, extraction discipline, and the fusion pipeline — follows an established curriculum, and the steps that borrow from it carry a link to the source at the point they borrow. Full references are at the end of step 12.

How to read the source labels

Every step is labeled with where its content comes from, so you always know whether you are being told something or shown a measurement.