🧠 graphify-in-action

A measured case study: 82–90% less context to answer a question — by querying a knowledge graph instead of re-reading the relevant source files.

Every number here was measured on a real project (2026-08-16, graphify 0.9.43). No estimates, no marketing — just data. 🧪

Read this in other languages: 🇹🇷 Türkçe

---

🌐 What is graphify?

graphify reads your code and draws a map of it — showing which files call, depend on, or affect each other. Instead of reading files one by one (like walking every street), your AI assistant just looks at the map and goes straight to what matters.

The FastAPI codebase mapped by graphify — every node is a concept, colors are communities. (Image: graphify, Apache-2.0)

• A map, not a pile of files. It pre-draws "who calls whom," so relationships are already there — the agent doesn't re-derive them every time.
• Read the map, not every street. One query returns a small, relevant slice (1,694 tokens) instead of re-reading the files an agent would otherwise grep and read (~9.6K–16.8K tokens).
• Local, free, offline. Built on your machine with 0 API calls — no keys, no data leaving your computer.

graphify path "FastAPI" "ModelField" — each hop is a call/import edge, so "who calls whom" is answered directly. (Image: graphify, Apache-2.0)

graphify is open-source (Graphify-Labs/graphify, 100k+ ★, YC) and works in 20+ assistants (Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, OpenCode, Aider, …).

📦 Official package: graphifyy on PyPI (CLI: graphify) · https://graphify.com · This repo is an independent case study, not affiliated with graphify.

TL;DR

• One graphify query returned a 1,694-token relationship map. Answering the same question without graphify means grepping and reading the relevant files — 9,636–16,842 tokens (5.7×–10× more).
• Extraction is 100% deterministic (tree-sitter AST, no LLM) — 0 API calls, fully offline.
• The graph also learns: 12 saved queries → 11 useful, 1 dead-end flagged so it is never repeated.

🚀 Measured impact (2026-08-16, graphify 0.9.43)

Metric	Value	Meaning
📉 Context reduction	82–90%	Same question: 1,694 tokens (graphify) vs 9,636–16,842 tokens (files an agent reads without graphify)
✅ Extraction quality	100% extracted · 0% ambiguous	4/931 edges inferred (0.4%) — deterministic, no LLM
💸 Cost	0 API calls · $0	tree-sitter local parse, fully offline
🗺️ Graph size	562 nodes · 931 edges · 44 communities · 599 KB	The whole code map in one file
🧠 Memory	12 queries → 11 useful · 1 dead-end	The system learns, dead-ends aren't re-scanned
⏱️ Answer time	~2 min	Agent reaches the right files via a scoped query

🎯 "Install it, but only if it helps" — answered with data

The install criterion was: "if graphify doesn't help, or worsens token/context usage, don't install it." Here is the measured answer:

Criterion	Measurement	Result
💨 Efficiency	9,636–16,842 tokens (files an agent reads without graphify) → 1,694 tokens per query	82–90% less context
🎯 Error-free	100% extracted · 0% ambiguous; inferred edges 0.4% (confidence 0.8)	Deterministic — no LLM, no hallucination source
💸 Cost	0 API calls, no API key required for code	Fully offline
🧠 Learning	12 saved queries: 11 useful, 1 dead-end flagged	Dead-ends aren't re-scanned

🔬 vs answering the same question without graphify

	Without graphify (grep + read)	A graphify query
Context	9,636–16,842 tokens — the agent greps and reads the relevant source files to map relationships	1,694 tokens — relationships pre-computed, delivered as a scoped subgraph
Relationships	Derived from scratch every session by reading files	931 edges pre-computed, delivered by the query
Latency / drift	Slows down in large contexts, can drift to the wrong file	Focused subgraph → the right file, in seconds

📂 What's in this repo

	English	Türkçe
📊 All measurements + real flows	docs/en/02-real-impact.md	docs/tr/02-gercek-etki.md
🔧 Install & daily usage guide	docs/en/01-installation-guide.md	docs/tr/01-kurulum-rehberi.md
⚖️ Neutral comparison vs alternatives	docs/en/03-comparison.md	docs/tr/03-karsilastirma.md
🎨 Interactive summary page	Live page (TR/EN toggle) — source: index.html

🏅 graphify's own published benchmarks

Benchmark	Metric	graphify
LOCOMO (n=300)	recall@10	0.497 (mem0 0.048, supermemory 0.149)
LongMemEval-S (n=50)	QA accuracy	76% (tied with dense RAG)
Graph build	LLM credits	0

Source: graphify BENCHMARKS.md

🚀 Get started (30 seconds)

uv tool install graphifyy        # or: pipx install graphifyy
graphify install --project --strict

Then, in your AI assistant, run /graphify . — you get three files (graph.html, GRAPH_REPORT.md, graph.json).

For a one-step AI-agent install prompt (with the "install only if it helps" decision criterion baked in), see the installation guide.

---

Source project: a multi-model financial forecasting platform (private repo, not linked here). Measured with graphify 0.9.43 (current at time of writing: 0.9.45).