Building Knowledge Graphs

Come join us at Kiln and learn about the building blocks that underpin memory systems for AI agents (and lots else besides): knowledge graphs. We'll get you started going from a pile of documents to entities and relationships you can use.

The main talk should be about 45 minutes, with plenty of time after for questions, conversations, and playing with the data. We've got two hours together, so the rest is informal experimentation and troubleshooting. We'll have food and beverages.

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

The knowledge graph will be computed entirely on your machine.

Who it is for#

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

What you'll get#

We'll provide a starter pack with datasets, examples, scripts, and config: everything you need to tweak and recreate the knowledge graph for each of the datasets we'll be providing, so you can start testing it.

You can start with the supplied examples, change the config, and regenerate them. You don't need to bring your own data to try things out.

What to bring#

Just coming for the talk? No setup needed.

Want to try the examples? Laptops encouraged! Follow the starter-pack setup instructions for macOS or Linux. If you have Windows you'll need WSL2 to run the demos. Basic familiarity with a terminal will help with running the scripts and changing config.

Want to try your own data? Bringing your own datasets is also encouraged. Start with a small collection you'd like to explore. The supplied datasets are the starting point for the session; working with your own is an experiment, not a promise that you'll leave with a finished graph.

You can use this prompt with your agent of choice to prepare a dataset:

Collect the dataset I select into a raw/ directory, with a separate
plain-text or Markdown file for each document. Keep the document text
intact and use unique filenames. Don't extract entities or relationships.

Create sources.jsonl next to raw/, with one JSON object per document.
Include its file path relative to the dataset directory (raw/...),
title, original source URL or path, and the timestamp the document was
created, if known. Use null for unknown metadata; don't invent it.

Choose documents you're comfortable giving to that agent. Preparing the data with an agent is separate from computing the knowledge graph locally.