qmd（QMD Markdown 搜索）

qmd - Local Search for Knowledge Bases

Search indexed markdown collections using keyword, semantic, or hybrid search.

When to Use

• Searching project documentation or knowledge bases
• Finding related files before making changes
• Discovering context across many markdown files
• Answering questions that span multiple documents

Quick Start

# Keyword search (fast, BM25)
qmd search "your query" --json -n 10

# Find related files
qmd search "topic" --files -n 20

# Semantic search (requires embeddings)
qmd vsearch "conceptual question" -n 10

# Hybrid search with reranking (best quality, slower)
qmd query "natural language question" --json -n 10

# Get full document content
qmd get "path/to/file.md" --full

Search Modes

Workflows

Before answering a question

qmd search "relevant keywords" --json -n 10

Review results, then read specific files for full context.

After making changes (update cascade)

qmd search "topic you changed" --files -n 20

Review all related files and update any that reference the changed topic.

Writing content that needs project context

qmd search "positioning voice tone" --json -n 10
qmd search "topic for content" --json -n 10

Index Management

# Update text index (fast, ~1 sec)
qmd update

# Update embeddings (slow, ~3-4 min)
qmd embed

# Check index health
qmd status

# List collections
qmd collection list

Setup & Troubleshooting

If qmd is not installed or broken, run the doctor script:

bash ~/.claude/skills/qmd/scripts/doctor.sh

For full installation and automation setup, see setup/README.md.

For the complete command reference, see references/commands.md.

Tips

• Results below score 0.3 are noise - ignore them
• qmd search (BM25) is good enough for day-to-day use
• Only use vsearch when keywords don't match concepts
• After changes, always run --files query to catch ripple effects
• Combine with Read tool: qmd finds files, Read gets full content

Self-Learning

Read LEARNED.md before using this skill. It contains hard-won lessons about what works and what doesn't.

Update LEARNED.md when you discover:

• A query that returned bad results and what worked instead
• Search mode choice that was wrong (e.g. search missed something vsearch found)
• Index issues (stale embeddings, missing files, broken binary)
• Installation or automation gotchas
• CLAUDE.md patterns that worked or failed for enforcing context-first behavior

Consolidation (keep LEARNED.md under 50 lines): Before adding a new entry, check the file length. If it's over 50 lines:

1. Merge duplicate/overlapping entries into single proven patterns
2. Remove entries older than 3 months that haven't been reinforced
3. Drop one-off observations that never recurred
4. Move detailed historical context to LEARNED-archive.md if worth preserving
5. Keep only entries that would change behavior - if it's obvious, cut it