Skip to content

getting-started.md

Latest commit

 

History

History
250 lines (183 loc) · 4.6 KB

getting-started.md

File metadata and controls

250 lines (183 loc) · 4.6 KB
Raw

Getting Started

A complete walkthrough from installation to a working wiki with searchable pages and a concept graph.

1. Install

# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/geronimo-iia/llm-wiki/main/install.sh | bash

# Or via cargo
cargo install llm-wiki-engine

Verify:

llm-wiki --version

See installation.md for all options.

2. Create a Wiki

llm-wiki spaces create ~/wikis/research --name research

This creates:

~/wikis/research/
├── README.md
├── wiki.toml
├── schemas/
│   ├── base.json
│   ├── concept.json
│   ├── paper.json
│   └── ...
├── inbox/
├── raw/
└── wiki/

The first wiki becomes the default. Check:

llm-wiki spaces list

3. Create a Page

llm-wiki content new concepts/mixture-of-experts --name "Mixture of Experts"

This scaffolds wiki/concepts/mixture-of-experts.md with frontmatter and a body template based on the type. Add --type concept for a concept-specific body structure:

llm-wiki content new concepts/mixture-of-experts --name "Mixture of Experts" --type concept

Frontmatter:

---
title: "Mixture of Experts"
type: page
status: draft
last_updated: "2025-07-20"
---

Edit the file — change the type, add content:

---
title: "Mixture of Experts"
type: concept
status: active
summary: "Sparse routing of tokens to expert subnetworks."
read_when:
  - "Understanding MoE architecture"
tags: [moe, scaling, sparse]
last_updated: "2025-07-20"
---

## Overview

MoE routes tokens to sparse expert subnetworks, trading compute
efficiency for model capacity.

## Key Ideas

- Each token is routed to a subset of experts
- Gating network decides which experts to activate
- Scales model parameters without scaling compute linearly

4. Ingest

llm-wiki ingest concepts/mixture-of-experts.md

This validates frontmatter against the concept schema, indexes the page in tantivy, and commits to git.

5. Add a Source

Create a paper page that the concept references:

llm-wiki content new sources/switch-transformer --name "Switch Transformer"

Edit wiki/sources/switch-transformer.md:

---
title: "Switch Transformer"
type: paper
status: active
summary: "Switch Transformer scales to trillion parameters using sparse MoE."
concepts:
  - concepts/mixture-of-experts
tags: [moe, scaling]
last_updated: "2025-07-20"
---

## Key Claims

- Simplified MoE routing with a single expert per token
- Scales to 1.6T parameters with improved training stability

Now update the concept page to reference this source — add to its frontmatter:

sources:
  - sources/switch-transformer

Ingest both:

llm-wiki ingest wiki/

6. Search

llm-wiki search "mixture of experts"

Output:

slug:  concepts/mixture-of-experts
uri:   wiki://research/concepts/mixture-of-experts
title: Mixture of Experts
score: 0.94

slug:  sources/switch-transformer
uri:   wiki://research/sources/switch-transformer
title: Switch Transformer
score: 0.81

Filter by type:

llm-wiki search "MoE" --type concept
llm-wiki search "MoE" --type paper

7. List Pages

llm-wiki list
concepts/mixture-of-experts      concept          active   Mixture of Experts
sources/switch-transformer       paper            active   Switch Transformer

Page 1/1 (2 total)

Filter:

llm-wiki list --type concept
llm-wiki list --status draft

8. View the Graph

llm-wiki graph
graph LR
  concepts_mixture_of_experts["Mixture of Experts"]:::concept
  sources_switch_transformer["Switch Transformer"]:::paper

  concepts_mixture_of_experts -->|fed-by| sources_switch_transformer
  sources_switch_transformer -->|informs| concepts_mixture_of_experts

  classDef concept fill:#cce5ff
  classDef paper fill:#d4edda
Loading

Filter by relation:

llm-wiki graph --relation fed-by
llm-wiki graph --root concepts/mixture-of-experts --depth 2

9. Connect an IDE

Start the MCP server:

llm-wiki serve

Add --watch for live indexing — external edits are picked up automatically:

llm-wiki serve --watch

Connect your editor — see ide-integration.md. Now the agent can search, read, write, and ingest through the wiki tools.

Next Steps