> Full SurrealDB documentation index: https://surrealdb.com/docs/llms.txt

# How the Robin Hood graph was built

The pipeline behind the Robin Hood demo: chapter splitting, serial ingest with a narrative clock, harvesting facts by chapter, alias suggestions and the three pages.

This page describes the pipeline that produced the [Robin Hood character graph](/docs/agent-memory/cookbooks/showcase/robin-hood.md), in the order the steps run. Each step is a short script, and the same steps work for any book or other text that arrives in parts. [Build your own character graph](/docs/agent-memory/cookbooks/showcase/robin-hood/build-your-own.md) gives the API calls to reproduce it.

## The corpus

The text is Project Gutenberg's [#10148](https://www.gutenberg.org/ebooks/10148): 22 chapters and about 107,000 words, and public domain, so the pages can ship with it. A full ingest takes about three hours, which is short enough to repeat after a change and compare the results.

## 1. Split the book into chapters

A script reads the Gutenberg HTML and writes one clean text file per chapter, split on the chapter headings.

## 2. Ingest the chapters in order, with a narrative clock

Each chapter is uploaded as a document and stamps it with `observedAt`, the time the memory treats as when the facts were learned. Chapter 1 is dated 1 January 2000, chapter 2 the next day, and so on, so the book reads as one day per chapter, and a reading position in the demo is exactly that date passed as `asOf`.

The chapters go in one at a time, in order. Extraction is shown the entities the Context already holds and told to reuse their names, so the order decides which form of a name later mentions attach to. After each chapter reports `ready`, the upload waits a little longer, because reconciliation lands shortly afterwards, and a chapter that starts before its predecessor's entities are committed mints fresh name variants instead of reusing them.

> [!IMPORTANT]
> Before a long ingest, upload one chapter and confirm that `GET /entities` is not empty. If extraction is not running, every later chapter still reports `ready` with nothing extracted, and the problem only shows at the end.

## 3. Harvest the graph by chapter

The next step reads every fact in the Context and groups them by the time they became known, which `observedAt` set to narrative time. That gives, for each entity, what was known about it at each reading position, and a dot's size on the graph is the number of facts known so far. The data comes from each entity's history, `GET /entities/{type}/{name}/history`, which returns every value with the time it became known and, for a replaced value, the time it stopped being current.

## 4. Describe each character from the facts

A script asks Gemini 2.5 Flash for a short description of each entity for each chapter, from the harvested facts and nothing else. Agent Memory plays no part in this step beyond supplying the facts. A description can only say what the memory captured, which is what makes it a fair picture of the memory rather than of the book.

## 5. Suggest links between names

A script suggests which names belong to one person, using two rules on entities of the same type:

| Rule | Example |
| --- | --- |
| The names differ only by an article | `host` and `the host` |
| Every capitalised word of the shorter name is in the longer one | `Tuck` and `Friar Tuck`, `merry Robin` and `Robin Hood` |

A name that fits two unrelated longer names is left unlinked. Each group takes the name with the most capitalised words as its main name, and the suggestions go into the page's data with the rule that produced each one.

## 6. Build the three pages

| Page | Content comes from |
| --- | --- |
| Graph | The harvested graph, descriptions and alias suggestions |
| Introduction | Written text, with every number filled in from the book and the graph |
| What the memory returns | The demo's own API calls, recorded with their responses and timings |

Each page is a single HTML file with no external dependencies, so it can be opened directly or hosted anywhere.