Main idea

Graphify, a tool that turns a code repository into a knowledge graph, chains two steps: a build, where an AI reads every file to infer the elements and their relations, then an exploration — splitting into groups, interactive map, report, path finding. The second step does not read the files: it reads a graph in the standard format of the NetworkX library, and does not check where it came from.

That indifference to provenance cuts the tool into two separable pieces. You can keep the exploration and replace the build with any producer of the same format: a 160-line script that reads the internal links of an Obsidian wiki is enough to feed the whole downstream part, with no API key.

The insertion point is not chosen at random: it is the exchange format between the steps, and that is precisely where the most expensive step hands over its result.

Why it matters

This gives a way to read a tool before adopting or rejecting it: look for the format that flows between its steps. A tool designed for another purpose becomes reusable as soon as you know how to produce, by another route, what its upstream step used to produce.

It also shifts the question of cost. What you pay for in such a tool is not the whole tool, it is the step you do not know how to replace.

Nuances and limits

A format that is accepted is not a format that is understood. The tool still expects what its original step used to produce: on notes, it raises 12,884 warnings because the links have no source file, a notion that only makes sense for code. The repurposing works, but alongside the tool's implicit contract, and an update can break it without warning.

Open questions

• Does a downstream step that has never been tested on a graph from elsewhere produce wrong results without flagging them, or only warnings?