Knowledge

Explore the Golden Suite knowledge graph — 2D map, full-text search, and per-entity notes linked across repositories and concepts.

Knowledge

The Knowledge section surfaces everything Golden has learned from public data — open-source repositories, research papers, Kaggle datasets, benchmark results — organized as a browsable graph.

Three surfaces, one underlying graph:

SurfacePathPurpose
Map/knowledge2D scatter of every entity, zoomable down to individual repos
Entity pages/knowledge/[entityId]Per-entity notes, k-NN neighbors, backlinks
Search/knowledge?q=...Hybrid full-text + semantic search across all entities

Data model

Every node in the graph is an entity — a repository, paper, dataset, or concept with a stable ID. Entities carry:

  • Notes — Obsidian-flavored Markdown that may contain wikilinks to other entities
  • Embedding — 768-dim vector used for map layout and nearest-neighbor lookup
  • Cluster assignment — produced by HDBSCAN at multiple zoom tiers for progressive disclosure
  • Source URL — canonical external link (GitHub, arXiv, Kaggle, etc.)

Regenerating the map

The static 2D positions are baked into frontend/public/knowledge-map.json and rebuilt from backend embeddings with UMAP + HDBSCAN + k-NN.

uv run --with numpy --with umap-learn --with python-dotenv \
       --with httpx --with "scikit-learn<1.6" --with "hdbscan==0.8.40" \
       scripts/build-knowledge-map.py

Re-run whenever you ingest a new batch of entities. The map component uses the 4-tier cluster_levels[] array to cross-fade labels as the user zooms.

Related

Knowledge Map →