0:00This video is about connecting documents to a knowledge graph. We'll look at the two links that join prose to structure, and then at how a vector database and a graph work together on one question.
0:13Real knowledge is not all in one place. Some of it sits in records with structure, and a lot of it sits in prose with none.
0:20This is a supplier quality report. It mentions Volt, and it has no structure at all.
0:26And this is Volt in the graph, which has structure, and which knows nothing about quality reports.
0:33Two bridges connect them. The first is extraction. Pull structure out of the text and put it in the graph as entities and relationships.
0:42The second is attachment. Keep the document itself linked to the entity it is about. We are going to do both, and then see why you need both.
0:52Let's watch extraction work on three documents that all mention the same supplier.
0:57The purchase record says Globex buys from Volt Limited, with Limited spelled as L T D. So extraction creates a node for Volt L T D.
1:06The location record spells it Volt Limited - this time Limited is spelled out. Different string, so extraction creates a second node.
1:15And the quality report writes the name Volt in capitals. That makes a third.
1:20Three identities for one supplier. The purchase fact and the location fact now hang off different nodes, so nothing connects them.
1:28And similar names never established identity on their own. Establishing it takes evidence, and it takes somebody making a decision.
1:36Extraction cannot fix this on its own. What fixes it is a decision.
1:41The three provisional identities get compared, and the decision is that they are one supplier. The canonical I-D is S seven.
1:50So S seven takes the place of the first provisional node. And the record behind it stops being a mention and becomes a source record, because it is now evidence for an identity somebody assigned.
2:02The Volt Limited identity consolidates into it.
2:05And its evidence reattaches. The location record now supports S seven.
2:10Same for the capitalised one.
2:12And the quality report reattaches as well.
2:14Now the purchase fact and the location fact meet at one identity, which is the thing that was broken.
2:22And every source record is exactly where it was. Nothing was deleted, so the decision can be looked at, and changed later if it turns out to be wrong.
2:33That is extraction done. Now the second bridge.
2:35Here is S seven on its own, with the relationships extraction gave it.
2:40And here are two of the documents it came from, still in their original wording, still under their own keys.
2:46Keep those linked to the entity rather than replacing them with edges. Extraction is lossy by design. It takes the facts it knows how to represent and it drops everything else.
2:59The prose holds the nuance, the exceptions, and the sentence somebody will want to read when they ask why the graph says what it says.
3:08Now the two ways people actually combine a vector database and a knowledge graph. Here is the first.
3:15Start with an ordinary similarity search, for battery supplier quality problems. Nothing special about that step at all.
3:22It comes back with a quality report. And that chunk carries a piece of metadata: the entity I-D it was resolved to.
3:30That I-D is what gets handed on. Not the supplier's name, not a string that looks like it. S seven, the identifier the graph itself uses.
3:39Which means the graph knows exactly where to start. Similarity picked the evidence, and the identity mapping turned that evidence into a starting point.
3:49If you want to see the retrieval and the expansion actually wired together, this is roughly what it looks like.
3:55Connect to the graph, then run the similarity search. The identifiers come out of the chunks' metadata, which is where the resolution work ends up living.
4:04That is the whole pattern. Retrieved identifiers become connected context.
4:09And the second half asks the graph for the neighbours of those identifiers. One query, matching on I-D rather than on name.
4:17Two honest caveats. The vector store call stands in for whichever one you already run, so there is no universal A-P-I there. And none of this works unless the chunks already carry resolved entity I-Ds.
4:31So the graph starts at S seven and looks at what is next to it.
4:35There is the entry point, and it is the only thing similarity gave us.
4:39One hop out, Volt is based in Northland.
4:42And two more. Globex and Initech both buy from Volt. Neither of those was in the retrieved chunks, and neither of them says anything about battery quality.
4:52One thing to be careful about. Expansion walked those purchase edges backwards to reach the customers, and that is allowed. But the edges still point the way they were recorded. Customers buy from Volt, not the other way round.
5:07And a second one. Being next to a supplier with a quality problem is not the same as having a quality problem. Expansion gives you context, and context still has to be read.
5:19The other combination runs the other way round.
5:22Start from the customer instead of from the query. Globex is C two, and we want documents that belong to Globex.
5:30Customer links or metadata give us three eligible documents. No scores yet, and that is deliberate. We are defining the boundary before ranking anything inside it.
5:41Now rank, and now the query matters. We are looking for late shipment causes.
5:46The delivery review moves to the top because its meaning is closest to that query, and the renewal notes drop to the bottom. Same three documents, different order.
5:57Structure decided which documents were eligible and similarity decided the order within them. The ranking never leaves the set scoping defined, which is what stops a semantic search wandering into another customer's paperwork.
6:12So here's where that leaves you.
6:14You can bridge text and graph in both directions. Extraction pulls entities and relationships out of prose, resolution makes them one identity, and the documents stay attached as evidence.
6:25And you can combine the two systems. Similarity finds an entry point and the graph expands from it, or structure scopes the documents and similarity ranks inside that scope.
6:37One thing to carry out of this. Neither system replaces the other, and the identifier on a chunk is what lets the two hand off to each other.

Documents and the Knowledge Graph, Together

Created by Abhyuday DesaiPublished 3 Oct, v1Runs 6 min 51 sec

0 views

About this video

Real knowledge sits partly in structured records and partly in prose. This video pulls structure out of three documents, watches extraction split one supplier into three identities, resolves them into one, and keeps the documents attached to the entity because extraction is lossy by design. Then the two combinations: a similarity search hands the graph an entity id to expand from, or the graph scopes which documents are eligible and similarity ranks inside that scope.

You'll learn how to:

Extract entities and relationships from text, and resolve the identities extraction splits
Keep source documents linked to the entity rather than replacing them with edges
Enter by similarity and expand the neighborhood by structure, with identifiers doing the handoff
Scope by structure first so a semantic search never wanders into another customer's paperwork

For developers combining a vector store and a knowledge graph in one retrieval system.

What changed

v103 Oct 2026abhydesai6 min 51 sec21ba95b

First published version.

0 views0 likes on this version