A 14-part technical course

Knowledge Graphs for AI Agents
From RAG Limits to Self-Updating Codebase Brains

Your coding agent reads 147 files to fix a 3-line bug because it has no map. This course teaches the three context layers that fix it — structural graphs (tree-sitter, Graphify, CodeGraph), narrative knowledge (Google's Open Knowledge Format), and vector RAG where it still belongs — plus the pipelines that keep them honest in production.

Start with Module 0 →

📂 The hands-on labs and the orderflow sample repo live in the course repository: github.com/varasrinivas/knowledge-graph-course — git clone https://github.com/varasrinivas/knowledge-graph-course.git

▶ Walkthroughs — four scenarios stepped through: the 147-file bug fix, a Graphify run, the MCP handshake, and the four-arm benchmark

13+1
modules + capstone
6
tracks
10
hands-on labs
1
mock monorepo you graph end-to-end

Structural

What the code is — deterministic AST call graphs. tree-sitter, Graphify, CodeGraph, okf-rs. Zero LLM cost, zero hallucination.

Narrative

What engineers know — curated Markdown + YAML concepts in Google's Open Knowledge Format. Git-native, agent-readable, one required field.

Semantic

The messy long tail — vector RAG over tickets, threads, and old docs. Still the right tool, once it stops being the only tool.

Track 0 — Overview

1 module
M00 ¡ Module 1 of 14

The Context Problem & the Knowledge Lifecycle

The 147-file bug fix, exploration vs reasoning tokens, Karpathy's LLM Wiki, the 90-day ecosystem, and the extract → curate → serve → maintain lifecycle. Code-free gateway.

~30 min ¡ code-free

Track 1 — Foundations

2 modules
M01 ¡ Module 2 of 14

Knowledge Graph Fundamentals

Nodes, typed edges, traversal vs search, communities, god nodes, blast radius — and the EXTRACTED / INFERRED / AMBIGUOUS honesty tags. Build a graph in pure Python.

~45 min ¡ lab M01
 M02 ¡ Module 3 of 14

The RAG Baseline and Where It Breaks

Shredded context, probabilistic retrieval vs canonical truth, the ACL leak, and where vector RAG still wins. Watch your own pipeline retrieve a stale definition.

~45 min ¡ lab M02
 M02B ¡ Bridge

The Context Engineering Frame

The coupling with the sibling course's M03B: six context layers, the four levers (add / compress / retrieve / offload) mapped onto this course's machinery, and context rot at two altitudes. Fix the poisoned transcript the two ways M03B's lab left open.

~40 min ¡ lab M02B

Track 2 — Structural Graphs

3 modules
M03 ¡ Module 4 of 14

Parsing Code into Graphs with tree-sitter

Fault-tolerant ASTs, extraction queries, deterministic zero-LLM parsing, and the name-resolution ladder from heuristics to LSP.

~60 min ¡ lab M03
 M04 ¡ Module 5 of 14

Graphify in Practice

/graphify end-to-end: three artifacts, Leiden communities, god nodes, path queries — and the honest read of the 71.5x benchmark (replications: 6.8x–49x).

~60 min ¡ lab M04
 M05 ¡ Module 6 of 14

The Tool Landscape: CodeGraph, okf-rs, and Friends

SQLite vs Markdown substrates, three-layer auto-sync, framework route recognition, per-language benchmark spread, and the comparison axes for any tool.

~45 min

Track 3 — Narrative Knowledge

3 modules
M06 ¡ Module 7 of 14

The Open Knowledge Format Specification

Bundles, concepts, one required field, index.md and log.md, progressive disclosure, and leniency by design — the one-page spec hiding a systems problem.

~60 min ¡ lab M06
 M07 ¡ Module 8 of 14

OKF for Codebases

From BigQuery tables to services: Responsibilities/Dependencies concepts, cross-links as a curated dependency graph, and the okf CLI's 13-rule lint.

~45 min ¡ lab M07
 M08 ¡ Module 9 of 14

The Enrichment Pipeline: Self-Updating Graphs

The two-pass agent (draft, then cite), diff-scoped scans as the entire cost model, lint gates, and the git hook that keeps the bundle honest.

~60 min ¡ lab M08

Track 4 — Serving & Architecture

2 modules
M09 ¡ Module 10 of 14

Serving Graphs to Agents over MCP

graph_callers in one tool call, the composite explore, 6,000-vs-15-token arithmetic, and the startup-cache trap.

~60 min ¡ lab M09
 M10 ¡ Module 11 of 14

The Hybrid Context Stack

Three layers, not three competitors: failure-mode-driven adoption, the staged query pipeline, the router pattern, and the ERPNext 70.8% → 82.0% case study.

~45 min

Track 5 — Production

2 modules
M11 ¡ Module 12 of 14

Drift, Staleness & Observability

Graphs don't self-heal: four documented silent hook failures, artifact desync, the O(V×E) scale wall, and the freshness gate that catches a lying map within 24 hours.

~60 min ¡ lab M11
 M12 ¡ Module 13 of 14

Structuring Projects So Agents Don't Get Lost + What's Next

Six project-structure patterns (layered CLAUDE.md, skills, subagents, deny rules, sparse worktrees, LSP) — progressive disclosure by hand — plus the road to OKF v0.2.

~45 min

Capstones

1 required ¡ 1 bonus
CAPSTONE 1 · Module 14 of 14 · ★★★★☆

Build a Self-Updating Knowledge Graph Pipeline

Wire everything together over the orderflow monorepo: structural graph + OKF bundle + enrichment hook + MCP server + CI freshness gate — then measure your real improvement multiple and explain it honestly.

7 phases ¡ 18 steps ¡ labs/CAPSTONE-knowledge-pipeline
 CAPSTONE 2 · Bonus · ★★★☆☆

Large Java Codebase — Measure the Token Reduction

Graph commons-lang (624 files) and spring-framework (9,195 files) with tree-sitter-java, serve them over MCP, and produce your own benchmark table: 33.7x per-question on spring, 567 files of exploration replaced by one tool call — with the honest per-question vs session-level framing.

6 phases ¡ real measured numbers ¡ labs/CAPSTONE-2-java-token-benchmark

Suggested path

  1. Everyone starts at M00 — the whole picture before the pieces.
  2. Tracks 1–3 in order: each module adds exactly one layer onto the orderflow running example.
  3. Already run agents in production? You may jump from M04 to M08 → M09 → M11, then return for M05/M10.
  4. Finish with the capstone — it reuses every lab solution you built along the way.