How it's built — archive · v0.1 · 2026-11-11

How AI reads and connects 15 years of work files

■ Method

Talk-style static page. Shows aggregate and verification numbers as charts and demos. No server or outbound API calls
Hero: three.js 0.160.0 Points + BufferGeometry for 11,905 document points (UMAP x·y, year as z). OrbitControls · theme filter · hover tip
Charts and demos: hand-built SVG/canvas. Embedding similarity, meaning map, RRF, extract, and version grouping replay measured or example values in the browser only
No training on this page. The pipeline it describes already ran on the office server (NVIDIA DGX Spark) with local models and code
LLM (this page): none. The original job used gemma4:26b for labeling/OCR/extract and bge-m3 for embeddings (local Ollama)
Fonts Instrument Serif · Pretendard (pinned CDN). No run ID (static publish)

Source: aggregates from one researcher's work SSD (2011–2026). Public page shows theme, type, year, and totals only — no source text, filenames, or internal titles · own data
Public totals: 20,010 originals · 548.5GB · 12,434 embedded docs · 11,905 map points · data/archive-data.json · data/archive-demo.json (measured)
Nature: measured aggregates plus demo examples (fictional docs, filenames, and RRF ranks labeled as examples)
No user input stored

Measurement: human spot checks from the original pipeline (its progress log). This page shows those numbers as-is
Text extract 12,519/12,549 (~99.8%) · versions 30/30 · labeling 8/10→10/10 · link table 6/10 (still weak) · on-theme in top-10 semantic hits (factory fire 10, BIM training 9, underground 8, biosignal 4–5)
Page checks: Node tests (facts, privacy) · Playwright (demos, present) · link checker. Train/eval split: not applicable

Public talk page and local aggregate numbers. Private-network search UI and full-text search are not on this site
Map points: embedding → UMAP (cosine, n_neighbors 30, min_dist 0.08, random_state 42). Positions are relative layout