23bfd15b19
Graph topology (Metatron's Cube pattern — 1 center + 12 outer nodes): - index: now links to ALL 12 other nodes (complete hub) - Inner ring cycle: hot→log→overview→dashboard→concepts/_index→entities/_index - Added: hot↔WikiMap, log↔sources/_index, dashboard↔concepts/_index - Added: entities/_index↔hot, entities/_index↔LLM Wiki Pattern - Added: sources/_index↔log, sources/_index↔entities/_index - Outer ring: concepts connected in triangle + Karpathy/sources cross-linked - Added: dashboard↔Compounding, entities/_index↔LLM Wiki Pattern graph.json physics for geometric arrangement: - repelStrength: 80 (strong push-apart for uniform spacing) - linkStrength: 3.0 (locks ring geometry) - linkDistance: 80 (tighter rings) - centerStrength: 0.25 (moderate center pull) - nodeSizeMultiplier: 2.0 (hub nodes visually dominant) - Added colors: questions=yellow, comparisons=red, nav=teal Phantom links removed from Hot Cache.md: - Removed [[Page A]], [[Page B]], [[New Page 1]], [[Existing Page]] New module pages: - wiki/questions/How does the LLM Wiki pattern work.md - wiki/comparisons/Wiki vs RAG.md - Adds questions/ and comparisons/ domains to the graph (yellow + red nodes)
2.1 KiB
2.1 KiB
type, title, subjects, dimensions, verdict, created, updated, tags, status, related, sources
| type | title | subjects | dimensions | verdict | created | updated | tags | status | related | sources | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| comparison | Wiki vs RAG |
|
|
Wiki wins at <1000 pages. RAG wins at enterprise scale. | 2026-04-07 | 2026-04-07 |
|
mature |
|
Wiki vs RAG
Overview
Both approaches let you query a large document collection. They differ fundamentally in when synthesis happens.
Comparison
| Dimension | LLM Wiki | Semantic RAG |
|---|---|---|
| How knowledge is stored | Pre-compiled markdown pages with cross-references already built | Raw chunks in a vector database |
| Finding answers | Read index → follow links → synthesize | Embed query → similarity search → assemble |
| Query cost | Low — synthesis already done | Higher — re-derives on every query |
| Infrastructure | Just markdown files | Embedding model + vector DB + chunking pipeline |
| Maintenance | Run a lint pass | Re-embed when content changes |
| Scale limit | ~hundreds of pages (index file navigation) | Millions of documents |
| Setup time | 5 minutes | Hours to days |
| Contradiction detection | Built in — LLM flags on ingest | Manual |
Verdict
Under 1000 pages → LLM Wiki. The index file is sufficient for navigation, token cost is low, setup is minimal, and the pre-compiled synthesis means every query benefits from everything ever read.
Over 100K pages → RAG. The index file becomes too large to read, and embedding-based retrieval becomes more efficient than full-index scanning.
The sweet spot: run the wiki pattern for active research (where things are being added, synthesized, and connected), then export to a vector store if the collection grows beyond the index threshold.
(Source: LLM Wiki Pattern, Compounding Knowledge)