build-graph

name	build-graph
description	Build a knowledge graph from markdown files. Ingests documents, extracts entities and relations, resolves duplicates, and induces a typed schema. Use when users want to create a knowledge graph, index their documents, or extract structured data from markdown files.

Build Knowledge Graph

Overview

This skill builds a complete knowledge graph from markdown files in the current project directory. It uses the kgmd MCP tools to ingest, extract, resolve, and induce.

Tools

Use the kgmd MCP server tools: init_graph, ingest_documents, get_chunks_for_extraction, store_extractions, get_resolution_candidates, apply_merges, get_induction_stats, store_schema.

Workflow

Step 1: Initialize

Call init_graph() to create the .kgmd/ database. If it already exists, skip this step.

Step 2: Ingest

Call ingest_documents() to find and chunk all markdown files. Report the count of documents found and chunks created.

Step 3: Extract (iterative)

Loop until complete:

1. Call get_chunks_for_extraction(batch_size=10) to get the next batch
2. If empty, extraction is complete — move to Step 4
3. For each chunk, extract entities and relations:
   • Entities: People, organizations, projects, technologies, events, concepts
   • Types: Use specific types (Person, Organization, Project, Technology, Event, Concept) or create new ones if none fit
   • Names: Use canonical full names (e.g., "Brian Anderson" not "Brian")
   • Attributes: Notable properties (role, location, date, industry, etc.)
   • Relations: Directed edges with predicates (works_at, founded, uses, depends_on, presented_at, manages, etc.)
   • Confidence: 0.7-1.0 based on how explicit the relationship is in the text
4. Call store_extractions() with the batch results
5. Report progress: chunks processed, entities found so far

Step 4: Resolve duplicates

1. Call get_resolution_candidates() to find similar entities via embedding clustering
2. For each candidate cluster, review whether they refer to the same real-world entity
3. Confirm merges where entities are clearly the same (e.g., "B. Anderson" and "Brian Anderson")
4. Reject merges where entities are different (e.g., "John Smith" the CEO vs. "John Smith" the engineer)
5. Call apply_merges() with confirmed merges

Step 5: Induce schema

1. Call get_induction_stats() to see entity type counts, relation predicates, and attribute summaries
2. Generate a YAML schema describing:
   • Entity types with their common attributes
   • Relation types with subject/object type constraints
3. Call store_schema() with the YAML

Step 6: Report

Summarize the completed graph:

• Documents processed
• Total entities by type
• Total relations by predicate
• Key findings or interesting patterns