# LLM Wiki

An Obsidian plugin that turns raw notes, clippings, and chat exports into an interlinked, AI-maintained personal wiki. Pages get YAML frontmatter, confidence scores, memory tiers, and graph edges — and the plugin keeps them healthy over time.

Designed as the knowledge core of a self-hosted second brain. Pairs naturally with [signal-bridge](https://github.com/enduserlab/signal-bridge), [signal-inbox](https://github.com/enduserlab/signal-inbox), and [perplexity-bridge](https://github.com/enduserlab/perplexity-bridge) for capture, and with [canvas-knowledge-graph](https://github.com/enduserlab/canvas-renderer) for spatial exploration.

## How it works

LLM Wiki implements four operations against a pool of raw sources:

1. **Ingest** — reads files from a raw sources folder, asks Claude to extract a title, summary, body, entities, tags, and confidence, then writes a structured wiki page with rich frontmatter and `[[wiki links]]`.
2. **Query** — runs local keyword search across wiki pages, or hands ranked context to Claude for a synthesised answer with cited sources.
3. **Lint** — scans the wiki for orphans, broken links, stale confidence, low-confidence pages, duplicate entities, missing frontmatter, empty pages, and tier mismatches.
4. **Crystallize** — picks groups of related pages and asks Claude to merge them into a higher-tier consolidated page, marking contradictions and creating `supersedes` edges.

A separate **memory lifecycle** runs Ebbinghaus-style confidence decay, evaluates tier promotions/demotions, and flags stale pages.

## Memory tiers

Every page lives in one of four tiers, inspired by human memory consolidation:

| Tier         | What it means                                      |
|--------------|----------------------------------------------------|
| `working`    | Just ingested, not yet validated                   |
| `episodic`   | Confirmed once, still tied to its source           |
| `semantic`   | Corroborated, generalised, durable                 |
| `procedural` | High-confidence, multi-source, long-lived          |

Pages start in `working` and get promoted as confidence grows and corroborations accumulate. They get demoted when confidence decays below the tier floor.

## Setup

### Prerequisites

- Obsidian 1.5.0+
- An Anthropic API key ([console.anthropic.com](https://console.anthropic.com))

### Install

1. Copy the `llm-wiki` folder into your vault's `.obsidian/plugins/` directory
2. Run `npm install && npm run build` inside the plugin folder
3. Enable "LLM Wiki" in Obsidian Settings → Community Plugins
4. Open LLM Wiki settings and enter your API key
5. (Optional) Adjust raw sources, wiki, and schema folder paths

### Folder layout

```
your-vault/
├── _sources/        ← drop raw files here for ingest
├── _schema/
│   └── graph.json   ← knowledge graph (nodes + edges)
└── wiki/            ← structured wiki pages with frontmatter
```

All paths are configurable in settings.

## Frontmatter shape

After ingest, wiki pages look like this:

```yaml
---
title: Retrieval-augmented generation
aliases:
  - RAG
tags:
  - llm
  - architecture
tier: semantic
source_type: llm-ingest
source_path: _sources/karpathy-talk.md
entity_type: concept
created: 2026-04-12T20:00:00.000Z
modified: 2026-04-15T08:30:00.000Z
related:
  - "[[Vector Databases]]"
  - "[[Embeddings]]"
supersedes: []
confidence:
  value: 0.78
  last_evaluated: 2026-04-15T08:30:00.000Z
  source_date: 2026-04-12T20:00:00.000Z
  corroborations: 3
  disputed: false
---
```

## Commands

- **Ingest all raw sources** — process every file in the raw sources folder
- **Ingest current file** — process the active file
- **Search wiki (local)** — fast keyword search, no API calls
- **Ask wiki (Claude-powered)** — Claude synthesises an answer from top-matching pages
- **Lint wiki** — surface health issues across the wiki
- **Run memory lifecycle** — apply confidence decay and tier evaluation
- **Show crystallize candidates** — list groups of pages that could be merged
- **Show wiki stats** — counts of nodes, edges, and orphans

## Settings

| Setting               | Default                       | Description                                          |
|-----------------------|-------------------------------|------------------------------------------------------|
| API key               | —                             | Your Anthropic API key                               |
| Model                 | Sonnet 4                      | Model used for ingest, query, crystallize            |
| Max tokens            | 4096                          | Maximum tokens per response                          |
| Raw sources folder    | `_sources`                    | Where unprocessed files live                         |
| Wiki folder           | `wiki`                        | Where structured pages are written                   |
| Schema folder         | `_schema`                     | Where `graph.json` is stored                         |
| Auto-ingest           | On                            | Automatically ingest new files dropped into sources  |
| Lint on startup       | On                            | Run lint when the vault opens                        |
| Confidence decay rate | 0.02                          | Daily exponential decay rate                         |
| Stale days            | 30                            | Days before a page is flagged stale                  |
| Default entity type   | concept                       | Used when ingest can't infer one                     |

## Knowledge graph

The plugin maintains a JSON-encoded graph at `_schema/graph.json`:

- **Nodes** — one per wiki page, indexed by file path. Stores title, entity type, tier, and confidence.
- **Edges** — typed, directed relationships: `related-to`, `part-of`, `derived-from`, `contradicts`, `supersedes`, `supports`, `example-of`, `prerequisite`.

The graph is the source of truth for orphan detection, crystallize candidate selection, and supersedes resolution. Pair with the [Canvas Knowledge Graph](https://github.com/enduserlab/canvas-renderer) plugin to render it as a live Obsidian canvas.

## Development

```bash
cd llm-wiki
npm install
npm run dev    # watch mode — rebuilds on save
```

Symlink the plugin folder into your test vault's `.obsidian/plugins/` directory. Reload Obsidian (Ctrl+R / Cmd+R) after rebuilds.

## License

MIT