Repository Knowledge Graph
An automatically-extracted map of how TinyMind’s code, docs, and tests connect — built with graphify.
Open the interactive graph Raw graph.json
What this is
graphify read the entire repository — every cpp/ header, every example, every doc and test — and turned it into a knowledge graph: a network of nodes (functions, classes, template types, files, concepts) joined by edges (calls, imports, references, conceptual links). It then ran community detection to find the natural module boundaries, with no help from the folder layout.
The result is a navigable picture of the codebase’s actual structure — how the pieces depend on one another — rather than how the directories happen to be arranged.
| Metric | Value |
|---|---|
| Files scanned | 373 (225 code · 118 docs · 30 images) |
| Nodes | 9,156 |
| Edges | 12,432 |
| Communities | 1,228 |
Two extraction passes feed the graph: a structural pass (deterministic AST parsing of the C++ headers — 8,773 nodes) and a semantic pass (LLM reading the docs and example write-ups for named concepts and cross-references — 383 nodes).
How to read it
Open the interactive graph. Because the full graph has more than 5,000 nodes, the view is aggregated to communities: each bubble is a cluster, sized by how many nodes it holds, and the lines between bubbles are cross-community links.
A community is a clump of things that mostly connect to each other. graphify draws a circle around densely-wired groups — more links inside the circle than crossing out of it. Nobody labels these by hand; they fall out of the link structure. For example, cpp/sigmoid.hpp’s 246 table-sizing templates reference each other and almost nothing else, so they form one community (“Sigmoid LUT Sizing”).
The largest, hand-named communities map cleanly onto the library’s real subsystems:
- Activation LUT families — separate communities for cos / sin / exp / log / sigmoid / tanh table sizing (
cpp/*.hpp) and value tables (8-bit, 32-bit, 128-bit). - Neural network core — gradient & weight updaters, the neuron connection model, layer feed-forward, node-delta / Adam backprop, output/classifier layer (
cpp/neuralnet.hpp). - Int8 quantization family —
QConv2D,QPointwiseConv2D,QLSTM,QCfC,QLayerNorm,QAttention Softmax, affine calibration params. - Continuous-time & KAN — self-attention, reverse-mode autodiff (
revdual.hpp), KAN backprop policy and B-spline tests. - Tooling & tests — NN / quantization / KAN / LUT-accuracy unit-test suites, the PyTorch importer.
Communities labeled Community N (an integer) are real clusters that simply weren’t given a topic name — most are tiny 1–4 node fragments in the long tail.
What the graph revealed
Tracing the graph surfaced one finding worth keeping:
The Q-format / neural-network pipeline and the int8 affine quantization pipeline are fully independent subgraphs. They share exactly one seam — cpp/qbridge.hpp — which imports qformat.hpp on one side and qaffine.hpp on the other. Remove that single bridge header and the two worlds fall into disconnected components. This confirms the architectural design rule stated throughout the docs: “None of the existing layers, QValue, or NeuralNet<> change; quantized models are built from a parallel layer family.”
A caveat when reading the graph. Generic template type names (
ValueType,size_t,InputType) appear once per file in the AST pass, but the connectivity report groups them by label. That makes a shared typedef likeValueTypelook like a single high-degree hub spanning many communities, when it is really 58 separate per-file typedefs with no edges between them. Treat any “most connected” claim on a generic type name skeptically — the node’s source file disambiguates it.
Regenerating
The graph is regenerated from the source on every Pages deploy — the published graph above always matches master, so it can’t drift from the code. Only the curated community labels (graphify-out/.graphify_labels.json) are versioned; the structural graph (graph.json/graph.html) is a build artifact and is not committed.
To rebuild and explore it locally:
/graphify . # full rebuild
/graphify . --update # re-extract only changed files (uses the cached pass)
/graphify query "How does the int8 attention path work?"
After adding code that forms new clusters, re-run /graphify . to label them and commit the updated .graphify_labels.json so the published graph names the new communities.
For the full picture — what’s versioned vs regenerated, the CI pipeline, and the “hairball trap” — see Maintenance.