Official skill · AI practice

Knowledge graph from documents

v1.0.0 · Updated 4 November 2026

Extracts entities and relationships from documents into triples with source quotes and confidence, merged names, a CSV-ready table, and a Mermaid diagram. Use it when you want to map people, companies, and products across many documents.

Works in Claude, ChatGPT, and coding agents. On Team and Business, the official skill in your pod updates itself when we improve it.

When to use it

• You want to map people, companies, and products across many documents
• You need a CSV for a graph tool or a RAG system

What you give it

• Documents or pasted text
• Entity types and relationships that matter
• Existing list of canonical names (optional)

What you get

An entity table, a triple table with quotes and confidence, a Mermaid diagram, and open questions about duplicates.

How it works

1. Fix the entity types and predicates first Write them as a short list: types in singular nouns (Person, Company, Product, Project), predicates as lowercase verb phrases (works_at, uses, owns, reports_to). Use one predicate per meaning; "employed by" and "works at" both become works_at.
2. Extract triples sentence by sentence Only what a sentence states. Each triple: subject, predicate, object, the exact source quote (under 25 words), the document and location. Don't infer relationships from proximity ("mentioned in the same paragraph" is not a relationship).
3. Grade confidence High: stated directly ("Mia leads Project Orbit"). Medium: stated but hedged or indirect ("Mia seems to be running Orbit now"). Low: implied by context. Our default: keep low-confidence triples in the table but out of the diagram.
4. Merge duplicates into canonical names Group variants ("Acme", "ACME Corp.", "Acme Corporation") under one canonical name and record the aliases. Different things with the same name get a qualifier ("Mercury (project)", "Mercury (vendor)"). When unsure whether two names are the same entity, keep them separate and list the question.
5. Check for conflicts Two triples that can't both be true (Mia reports_to Jon; Mia reports_to Ana) are flagged with both sources and dates. Prefer the newer source only if the documents are dated; otherwise leave it open.
6. Write the outputs An entity table, a triple table in CSV-ready form (no merged cells, one row per triple), and a Mermaid flowchart of high and medium triples. In Mermaid, put labels with spaces or parentheses in double quotes, and never use the bare word "end" as a node id.

The SKILL.md

The whole file, as your AI reads it. Download it

---
name: knowledge-graph-extraction
description: Extracts entities and relationships from documents into a knowledge graph - entity types, subject-predicate-object triples with the source quote and a confidence level, duplicates merged under one canonical name - output as a CSV-ready table and a Mermaid diagram. Use when someone asks to map who works with whom, which customers use which products, turn documents into a knowledge graph, extract entities and relationships, build a graph from notes, or make data for GraphRAG or Neo4j.
---

# Knowledge graph extraction

A knowledge graph is a set of statements of the form subject–predicate–object ("Acme Corp — uses — Ledgerly"), the same three-part shape as a W3C RDF triple. It is only useful if each statement can be traced to a source sentence and the same thing always has the same name. So every triple here carries its quote and a confidence level, and duplicates are merged before anything is drawn.

## When to use

- Turning meeting notes, contracts, CRM notes, or reports into a map of people, companies, products, and how they relate.
- Preparing a CSV to import into a graph tool (Neo4j, Kumu, a spreadsheet pivot) or a RAG system.
- Answering "who is connected to what?" across many documents.
- Not for designing the classes and rules of a domain from scratch: use ontology-draft first if the entity types aren't known.

## What you need

- The documents, or pasted text. Number them (D1, D2...) so each triple can cite its source.
- The entity types and relationships that matter. If not given, propose 4 to 8 entity types and 5 to 10 predicates from a first read, mark them "[proposal]", and continue.
- Any existing list of canonical names (customer list, staff directory). If none, build one in step 4.

## Steps

1. **Fix the entity types and predicates first.** Write them as a short list: types in singular nouns (Person, Company, Product, Project), predicates as lowercase verb phrases (works_at, uses, owns, reports_to). Use one predicate per meaning; "employed by" and "works at" both become works_at.
2. **Extract triples sentence by sentence.** Only what a sentence states. Each triple: subject, predicate, object, the exact source quote (under 25 words), the document and location. Don't infer relationships from proximity ("mentioned in the same paragraph" is not a relationship).
3. **Grade confidence.** High: stated directly ("Mia leads Project Orbit"). Medium: stated but hedged or indirect ("Mia seems to be running Orbit now"). Low: implied by context. Our default: keep low-confidence triples in the table but out of the diagram.
4. **Merge duplicates into canonical names.** Group variants ("Acme", "ACME Corp.", "Acme Corporation") under one canonical name and record the aliases. Different things with the same name get a qualifier ("Mercury (project)", "Mercury (vendor)"). When unsure whether two names are the same entity, keep them separate and list the question.
5. **Check for conflicts.** Two triples that can't both be true (Mia reports_to Jon; Mia reports_to Ana) are flagged with both sources and dates. Prefer the newer source only if the documents are dated; otherwise leave it open.
6. **Write the outputs.** An entity table, a triple table in CSV-ready form (no merged cells, one row per triple), and a Mermaid flowchart of high and medium triples. In Mermaid, put labels with spaces or parentheses in double quotes, and never use the bare word "end" as a node id.

## Output format

````markdown
## Entities
| id | canonical_name | type | aliases |
|---|---|---|---|
| E1 | [name] | [type] | [alias; alias] |

## Triples
| subject | predicate | object | confidence | source_quote | source |
|---|---|---|---|---|---|
| [canonical name] | [predicate] | [canonical name] | high/medium/low | "[exact quote]" | D[n], [section/line] |

## Diagram (high and medium only)
```mermaid
flowchart LR
  E1["[name]"] -->|[predicate]| E2["[name]"]
```

## Open questions
- [Possible duplicate / conflict / unclear relationship], sources D[n], D[n]
````

## Example

**Input:**
D1 (call notes, 3 May): "Mia Chen from Brightwave said they've rolled out Ledgerly to their finance team. Their CFO, Jon Park, signed off."
D2 (email, 10 May): "BrightWave Inc. wants to add Ledgerly Reports. Mia mentioned Tessa might take over the project."

**Output:**

````markdown
## Entities
| id | canonical_name | type | aliases |
|---|---|---|---|
| E1 | Brightwave | Company | BrightWave Inc. |
| E2 | Mia Chen | Person | Mia |
| E3 | Jon Park | Person | |
| E4 | Ledgerly | Product | |
| E5 | Ledgerly Reports | Product | |
| E6 | Tessa | Person | |
| E7 | Ledgerly rollout | Project | |
| E8 | Ledgerly project | Project | |

## Triples
| subject | predicate | object | confidence | source_quote | source |
|---|---|---|---|---|---|
| Mia Chen | works_at | Brightwave | high | "Mia Chen from Brightwave" | D1, line 1 |
| Brightwave | uses | Ledgerly | high | "they've rolled out Ledgerly to their finance team" | D1, line 1 |
| Jon Park | works_at | Brightwave | high | "Their CFO, Jon Park" | D1, line 2 |
| Jon Park | approved | Ledgerly rollout | high | "Jon Park, signed off" | D1, line 2 |
| Brightwave | interested_in | Ledgerly Reports | high | "wants to add Ledgerly Reports" | D2, line 1 |
| Tessa | may_lead | Ledgerly project | medium | "Tessa might take over the project" | D2, line 2 |

## Diagram (high and medium only)
```mermaid
flowchart LR
  E2["Mia Chen"] -->|works_at| E1["Brightwave"]
  E3["Jon Park"] -->|works_at| E1
  E1 -->|uses| E4["Ledgerly"]
  E1 -->|interested_in| E5["Ledgerly Reports"]
  E3 -->|approved| E7["Ledgerly rollout"]
  E6["Tessa"] -->|may_lead| E8["Ledgerly project"]
```

## Open questions
- Does Tessa work at Brightwave or at our company? D2 doesn't say.
- "Ledgerly rollout" and "Ledgerly project" may be the same project; kept separate.
````

## Checks before you send

- [ ] Every triple has an exact quote that appears in the cited document.
- [ ] Each real-world entity appears under one canonical name, with aliases listed.
- [ ] Each meaning uses one predicate; no two predicates mean the same thing.
- [ ] The diagram contains only high and medium triples and renders as valid Mermaid.
- [ ] Possible duplicates and conflicts are listed as open questions, not silently merged.

## Common mistakes

- Treating co-occurrence as a relationship. Two names in one paragraph aren't linked unless a sentence says how.
- Letting the predicate list grow to 40 near-synonyms. Merge them before extracting.
- Drawing every triple. A diagram with more than about 30 edges is unreadable; split it by entity or type.

Share it with your team

Add it to an AI Skill Pod and everyone on the team, contractors included, uses the same version from their own AI. Read the best ways to share AI skills with a team

Add to my team's pod

Changelog

• v1.0.0 · 4 November 2026 · First release.