QMD Search

Search your project's indexed markdown documentation using qmd's MCP tools. This skill reads your project's qmd configuration to know which collections to search, constructs structured queries, and retrieves relevant documents.

Step 1: Read Config

Read .claude/qmd.json from the project root.

If missing, tell the user:

qmd is not configured for this project. Run /avinyc:qmd-configure to set it up.

Then STOP.

Extract project name and collections — each has a name, path, and description. Pick the best collection for the query by matching against descriptions. If only one collection exists, use it.

Step 2: Search via MCP

Use the mcp__qmd__query tool with a structured query. This is the primary search method — it's faster than CLI (models stay warm between queries) and supports multi-type queries with intent disambiguation.

Constructing the Query

Every search should include at least two query types for best recall. The first query gets 2x weight in fusion scoring, so put your best guess first.

{
  "searches": [
    { "type": "lex", "query": "2-5 exact keywords, no filler" },
    { "type": "vec", "query": "full natural language question" }
  ],
  "collections": ["<collection_name>"],
  "limit": 10
}

Query Types

Type	Method	When to use	Writing tips
lex	BM25 keyword	You know the exact terms in the docs	2-5 terms. Exact phrases: "rate limiter". Exclude: -sports. Code identifiers work.
vec	Vector semantic	You don't know the vocabulary	Full question. Be specific: "how does the auth system handle session expiry?"
hyde	Hypothetical doc	Complex topic, you can imagine the answer	Write 50-100 words of what the answer looks like, using vocabulary you expect in the result.

When to Use What

Situation	Query types to include
Know exact terms	lex only
Don't know vocabulary	vec only
Best recall	lex + vec
Complex or broad topic	lex + vec + hyde
Ambiguous query (e.g., "performance" could mean web perf, team health, etc.)	Any combination + intent

Intent Disambiguation

When a query term is ambiguous, add intent to steer all pipeline stages (expansion, reranking, snippet extraction):

{
  "searches": [
    { "type": "lex", "query": "performance" },
    { "type": "vec", "query": "how to improve page load speed" }
  ],
  "intent": "web page load times and Core Web Vitals",
  "collections": ["project_docs"],
  "limit": 10
}

Intent does not search on its own — it's a steering signal that disambiguates what you mean.

Step 3: Interpret Results

Results include docid, score, file, title, context, and snippet.

Score	Meaning
0.7+	Highly relevant — read this document
0.5–0.7	Worth reading if topic matches
< 0.5 on all	Try different query types or refine

Step 4: Retrieve Documents

To read a full document from the results, use mcp__qmd__get:

{ "file": "#docid" }

Or by file path:

{ "file": "collection_name/path/to/file.md" }

For multiple related documents, use mcp__qmd__multi_get to batch retrieve:

{ "pattern": "collection_name/docs/*.md" }

This is faster than reading files one at a time.

Error Handling

If MCP tools fail, diagnose the issue:

Error	Likely cause	Fix
mcp__qmd__query not available	MCP server not configured	Run claude mcp add qmd -- qmd mcp or add qmd to .mcp.json. Run /avinyc:qmd-configure.
MCP call returns error	Server not running or crashed	Run qmd mcp to verify. Check qmd status.
Collection not found	Config out of sync	Run /avinyc:qmd-doctor to diagnose.
No results	Index empty or stale	Run qmd update && qmd embed to rebuild.

CLI Fallback

If MCP tools are unavailable, fall back to the CLI. The qmd query command supports structured queries via multiline strings:

qmd query $'lex: exact keywords here\nvec: natural language question here' -c <collection> --json -n 10

This single command subsumes qmd search (BM25 only) and qmd vsearch (vector only). Always use --json for parseable output.

Fallback to Glob/Grep

After 2 poor query attempts (different query types, refined terms), fall back to Glob/Grep on the collection's directory path from config. qmd can't find everything — sometimes a direct file search is faster.