Outils — vue longue

Toutes les pages concaténées sur un seul document, pour un Ctrl-F direct.

qmd Search Engine

Local search engine for markdown files designed specifically for llm-wiki-pattern implementations. Provides hybrid BM25/vector search with LLM re-ranking, all running on-device. Essential tool for scaling wiki navigation beyond simple index-based browsing.

Core Capabilities

Hybrid Search Architecture:

BM25 for keyword-based relevance scoring
Vector embeddings for semantic similarity
LLM re-ranking for improved result quality
Combines best of lexical and semantic search

Local Operation:

Runs entirely on-device, no external services required
Privacy-preserving - content never leaves local machine
Fast response times for interactive exploration
No API costs or rate limits

Integration Options

CLI Interface:

LLM can shell out to qmd for search operations
Scriptable for automated wiki maintenance tasks
Batch processing capabilities for large-scale analysis

MCP Server:

Native tool integration for LLM agents
Structured query and response format
Seamless integration with agent workflows

Use Cases in LLM Wikis

Query Workflow Enhancement:

Replace simple index.md scanning with sophisticated search
Find relevant pages across large wiki collections
Support complex multi-term and semantic queries

Lint Workflow Support:

Find orphan pages and missing cross-references
Identify contradictions between related pages
Discover content gaps and redundancies

Exploration and Discovery:

Surface unexpected connections between concepts
Find related content across different categories
Support research and synthesis workflows

Scaling Considerations

When to Implement:

Essential when wiki grows beyond ~100 pages
Index.md becomes unwieldy for navigation
Complex queries require semantic understanding

Performance Characteristics:

Handles thousands of markdown files efficiently
Real-time search for interactive exploration
Batch analysis for maintenance tasks

Alternative Approaches

While qmd is recommended by andrej-karpathy, the llm-wiki-pattern is tool-agnostic. Simple alternatives include:

Naive Search Scripts:

Basic grep-based search over markdown files
LLM can help implement custom search logic
Sufficient for smaller wikis or specific domains

Existing Tools:

Standard text search utilities (ripgrep, ag)
Markdown-specific search tools
Platform-native search (macOS Spotlight, Windows Search)

Implementation Strategy

Incremental Adoption:

Start with index.md for small wikis
Add basic search when navigation becomes cumbersome
Implement qmd or similar when semantic search becomes valuable

LLM Integration:

Configure search tool in schema layer
Train LLM on search syntax and capabilities
Integrate search into standard workflows