Skip to content

graph-tools.md

Latest commit

 

History

History
119 lines (86 loc) · 3.34 KB

graph-tools.md

File metadata and controls

119 lines (86 loc) · 3.34 KB
Raw

Graph tools

I want to see the graph without depending on an external interface. NetworkX can inspect its structure, PyVis makes an interactive HTML page and Graphviz produces an SVG.

Install

Use the Python 3.12 LangGraph lab:

cd ~/knowledge-graphs/langgraph-lab
uv add 'networkx==3.7' 'pyvis==0.3.2' 'jupyterlab==4.6.4'

The NUC has graph packages inside the Hailo and other environments. Leave those with their applications; these versions go in the lab.

Find cycles and draw the graph

Save this as inspect_graph.py next to loop.py:

from pathlib import Path
import re

import networkx as nx
from pyvis.network import Network

from loop import graph

workflow = graph.get_graph()
g = nx.DiGraph()
g.add_nodes_from(workflow.nodes)
g.add_edges_from((edge.source, edge.target) for edge in workflow.edges)

print("Short cycles:", list(nx.simple_cycles(g, length_bound=6)))
print("Isolated nodes:", list(nx.isolates(g)))
print("Dead ends:", [n for n in g if n != "__end__" and g.out_degree(n) == 0])
print("Unreachable:", sorted(set(g) - nx.descendants(g, "__start__") - {"__start__"}))

Path("loop.mmd").write_text(workflow.draw_mermaid(), encoding="utf-8")
view = Network(height="750px", width="100%", directed=True, cdn_resources="in_line")
view.from_nx(g)
html = view.generate_html(notebook=False)
# PyVis adds remote Bootstrap files that this view does not need.
html = re.sub(r'<link\b[^>]*href="https?://[^"]+"[^>]*>', "", html)
html = re.sub(r'<script\b[^>]*src="https?://[^"]+"[^>]*>\s*</script>', "", html)
Path("loop.html").write_text(html, encoding="utf-8")

Run it:

uv run python inspect_graph.py

Open loop.html in the browser. The Mermaid version is in loop.mmd.

The generate -> validate -> generate cycle is expected. The loop is supposed to retry. NetworkX finds the cycle; running the loop tells me whether it stops.

Moving nodes around only changes the drawing. To fix the loop, edit loop.py, run it again and regenerate the HTML. For stored facts, use the Graphiti correction.

Graphviz

The NUC already has dot. Save this as loop.dot:

digraph loop {
  rankdir=LR;
  generate -> validate;
  validate -> generate [label="failed, retries left"];
  validate -> review [label="valid"];
  validate -> end [label="limit"];
  review -> end [label="approve or reject"];
}

Generate the image:

dot -Tsvg loop.dot -o loop.svg

This one is a hand-written diagram. Update it when the code changes. Graphviz supports other output formats too.

JupyterLab

Start it in the NUC lab:

uv run jupyter lab --ip=127.0.0.1 --port=18888 --no-browser

From the Mac:

ssh -N -o ExitOnForwardFailure=yes \
  -L 127.0.0.1:18888:127.0.0.1:18888 pink-sudo

Open the URL with the token that Jupyter prints. Keep its authentication enabled and the listener on localhost. A notebook can run commands as its user.