Workshop

Building Knowledge Graphs

From a pile of documents to entities and relationships you can query, computed on your own machine

Length: 2 hours
For: Anyone who has wanted to use the relationships in their data to answer questions that span documents. No graph database experience needed

Arrow keys advance. P shows the notes. Esc comes back here.

01
antfly.io/events/building-knowledge-graphs
02
What we cover
- Extracting entities and relationships from documents with a local model
- Storing relationships alongside the source documents themselves
- Querying across the graph and the text in one request
03
Contents
1. Why Graphs
2. From Documents to a Graph
3. Why Not Just Ask an Agent?
04
Why Graphs
05
AI search engines for science
- Often irrelevant results
- Pretty slow
- Bad precision
Notes
It was late 2024 and I wanted to search scientific articles. With AI getting pretty impressive, surely some new startups were blowing Google Scholar out of the water. One of the motivating problems for Antfly: searching scientific articles in the AI era should be better than Google Scholar, and better than "RAG over all science." The biggest problems with the products I tried: search was slow, completely irrelevant results were shown prominently, and key words were misinterpreted. The goal is useful evidence, not just more generated text.
06
We take our agents having access to web-scale data as a given
07
The OG web-scale graph: PageRank
"Everyone links to Website B, so it must be important."

"Website B links to Website C, so it probably is as well."
Notes
Links as a ranking signal, in about 20 seconds. This is how Google did it originally, and it tamed the Internet in a way nothing had before. It is a graph-ranking example, not the same thing as an extracted knowledge graph.

Image: "PageRanks-Example.svg" by 345Kai and Stannered, with percentage signs added by Frakturfreund. Public domain. https://commons.wikimedia.org/wiki/File:PageRanks-Example.svg
08
From Documents to a Graph
09
But what if you just have a pile of documents?

Notes
We want to extract things from documents and infer the connections within and between them.
10
11
Try it yourself
Ask your agent:
agent harness
Set up https://github.com/antflydb/knowledge-graphs-workshop
Or do it yourself:
terminal
git clone https://github.com/antflydb/knowledge-graphs-workshop cd knowledge-graphs-workshop make up # → http://127.0.0.1:8000 (Ctrl-C stops everything)
Prerequisites: uv and Antfly v0.3.0-rc.1 on your PATH (set ANTFLY_BIN if an older antfly is installed).
The first run pulls GLiNER2.5 and EmbeddingGemma (~2 GB) and the prebuilt database (~350 MB); after that it starts in seconds.
Notes
Slides and setup: antfly.io/learn/workshops/building-knowledge-graphs. make on its own lists every target.
12
Notes
What we want: pass in the types of things you want to extract and the relationships between them, and get a graph back.
13
Entity extraction from text
Entity types Relationship vocabulary
character parent_of
house member_of
kingdom rules
city located_in

Notes
You supply the vocabulary and the relationship definitions. These are the configured types and predicates from the ASoIaF dataset's config.yaml. They are a control knob, not a guarantee of correctness: note that Winterfell came out as a kingdom in one passage and a city in another.
14
15
Entity resolving

Notes
The next step beyond this workshop's pipeline. The workshop graph groups names by type and spelling ("Tyrion's" joins "Tyrion"), but "Tyrion" and "Tyrion Lannister" stay separate: it does not resolve aliases, surnames or pronouns.
16
Pass through a decision model

Try to weed out weak or untrue relations.

Ordered pair Saved decision
Eddard "Ned" Stark → Winterfell rules 0.575
"is the Lord of Winterfell" no_relation 0.241
unclear 0.184
Notes
Walk through the ordered pair, the source wording, the three candidate labels, and the recorded choice. The decision model is told to respect direction, negation, uncertainty and mixed continuity, and to use the supplied passage rather than outside knowledge. The score is a model output, not a calibrated probability that the relationship is true.
17
Change the config and rerun
Notes
Turn the errors you saw into small configuration experiments, and test one change at a time against saved examples. A schema edit is a hypothesis, not an automatic fix. After editing a dataset's config.yaml, make build rebuilds it.
18
Things to try

Search the ASoIaF graph:
- Why does Tyrion kill Tywin?
- Jon Snow mother
- Dorne
Then pick one relationship:
1. Open its supporting passage.
2. Does the passage support the relationship's type and direction?
3. Change one choice in config.yaml, rebuild, and compare.
19
Why Not Just Ask an Agent?
20
Notes
Why go through the trouble of building this when I could just ask an agent for the answer?
21
22
23
24
What Claude did
- Used a bunch of knowledge outside the text corpus
- Used 18 subagents and $62 of API credits
- Merged entities "by hand"
- Its agents used varying conventions
Agents can totally build knowledge graphs, but you should give them the right tools to do it.
25
Why not just ask an agent?
- Keep data on-prem
- More reproducible, scalable, and deterministic
- Control terminology
- Run it yourself, own it yourself
- An agent won't know your data. Hallucinations are definitely still a thing
26
Why else? Enterprise

Each of our systems is internally coherent and externally isolated. Atlas, our telemetry platform, knows applications and their time series. Radar, our alerting system, knows thresholds and firing history. incident.io knows severities and responders. PagerDuty knows rotations. Spinnaker knows what shipped and when. And our Slack history knows, in a form no query engine can reach, which engineer solved something very similar last quarter and what they found.
The difficulty is that they do not agree on what they are talking about.

Ontology-Driven Observability: Building an End-to-End Knowledge Graph at Netflix Scale netflixtechblog.com

Entity types	Relationship vocabulary
character	parent_of
house	member_of
kingdom	rules
city	located_in

Ordered pair	Saved decision
Eddard "Ned" Stark → Winterfell	rules	0.575
"is the Lord of Winterfell"	no_relation	0.241
	unclear	0.184

Other tools

Tool	What it's for
Graphiti	Real-time knowledge graphs for agent memory
GraphRAG	Graph-based retrieval-augmented generation, from Microsoft
LightRAG	Simple, fast graph-based RAG
LLM Graph Builder	Neo4j graphs from unstructured documents
Cognee	Open-source memory platform for agents
Mem0	A memory layer for agents and apps

28
Apollo 13
29
Thank you

antfly.io/learn/workshops/building-knowledge-graphs · github.com/antflydb/knowledge-graphs-workshop

Events

Building Knowledge Graphs
Tuesday, October 20, 2026 · Kiln, Portland

Building Knowledge Graphs

What we cover

Contents

Why Graphs

AI search engines for science

We take our agents having access to web-scale data as a given

The OG web-scale graph: PageRank

From Documents to a Graph

But what if you just have a pile of documents?

Try it yourself

Entity extraction from text

Entity resolving

Pass through a decision model

Change the config and rerun

Things to try

Why Not Just Ask an Agent?

What Claude did

Why not just ask an agent?

Why else? Enterprise

Other tools

Apollo 13

Thank you

Events