QMD - Quick Markdown Search

Overview

QMD is an on-device search engine for markdown-based knowledge bases. It provides three search modes with increasing quality/latency tradeoffs, plus document retrieval capabilities. All processing runs locally using embedded GGUF models.

Search Modes

Choose the appropriate search mode based on the query type:

1. BM25 Search (qmd search) - Fast Keyword Search

qmd search "<query>" -n 10 --json

When to use:

• Exact keyword or phrase searches
• Technical terms, function names, specific identifiers
• When speed matters more than semantic understanding

2. Vector Search (qmd vsearch) - Semantic Search

qmd vsearch "<query>" -n 10 --json

When to use:

• Conceptual queries ("how do I...", "what about...")
• When exact keywords are unknown
• Finding related content with different wording

3. Hybrid Search (qmd query) - Highest Quality

qmd query "<query>" -n 10 --json

When to use:

• Important searches where accuracy matters
• Complex queries requiring understanding
• Default choice when unsure which mode to use

Document Retrieval

After searching, retrieve full document content:

# By file path
qmd get "collection/path/to/file.md" --json

# By document ID (from search results)
qmd get "#a1b2c3" --json

# Multiple documents
qmd multi-get "collection/*.md" -l 500 --json

Typical Workflow

1. Search to find relevant documents:

   qmd query "quarterly planning" -n 5 --json

2. Review search results (docids, scores, snippets)

3. Retrieve full content of relevant documents:

   qmd get "#abc123" --json

4. Synthesize information from retrieved documents to answer the user's question

Common Options

Option	Description
-n <num>	Number of results (default: 5)
--json	JSON output for parsing
--files	TSV format: docid, score, filepath
-c <name>	Restrict to specific collection
--min-score <n>	Minimum relevance threshold (0-1)
-l <lines>	Max lines per document

Index Status

Check available collections and index health:

qmd status

Critical Constraints

Never run these commands automatically:

• qmd collection add - Expensive indexing operation
• qmd embed - Generates embeddings (slow, resource-intensive)
• qmd update - Re-indexes all collections

These operations should only be run when explicitly requested by the user.

Score Interpretation

Score Range	Meaning
0.8 - 1.0	Highly relevant
0.5 - 0.8	Moderately relevant
0.2 - 0.5	Somewhat relevant
0.0 - 0.2	Low relevance

Resources

references/

• cli_reference.md - Complete CLI API documentation with all commands, options, and examples