search

QMD Search Guide

Current State

• Collections: !qmd status 2>/dev/null | grep -c '^ ' || echo "0"
• Status: !qmd status 2>/dev/null | head -1 || echo "QMD not running"

When to use QMD

QMD searches indexed collections: external codebases, notes, docs you've added with qmd collection add. For files in the current working directory, use Grep and Glob instead. They're faster, cheaper on context, and more precise.

Modality

Approach	When to use	How
Typed queries (MCP + CLI)	You know what you need. Exact terms, semantic concepts, or both.	MCP: searches: [{type: "lex", query: "handleAuth"}, {type: "vec", query: "how auth works"}]. CLI: qmd query $'lex: handleAuth\nvec: how auth works'
Auto-expand (CLI only)	Most CLI queries. Let the pipeline decide.	CLI: qmd query "auth middleware". Not available via MCP.

Use query for everything. MCP accepts typed sub-queries (lex, vec, hyde). CLI also supports expand: and plain text (auto-expand).

How query Works

Hybrid pipeline with GRPO-optimized query expansion → BM25 + vector search → RRF fusion → LLM cross-encoder reranking → position-aware blending → dedup. For pipeline internals (sub-query types, blending weights, conditional expansion logic), see references/pipeline.md.

Documents are chunked into 900-token pieces with 15% overlap. Search matches point to the chunk, not the exact line. Use get with fromLine/maxLines to narrow down.

Query Document Format

The query tool accepts an array of typed sub-queries. MCP uses structured JSON: searches: [{type: "lex", query: "term"}, ...] with types lex, vec, hyde only. CLI uses a text format: each line has an optional type prefix (lex: term). CLI also supports expand: and plain text (auto-expand).

Grammar

MCP type enum: "lex" | "vec" | "hyde". CLI adds "expand" and implicit expand (plain text).

root    ::= line ("\n" line)*
line    ::= (type ":" SP)? content
type    ::= "lex" | "vec" | "hyde" | "expand"   # expand is CLI-only
content ::= [^\n]+

Types

Type	Routed to	Purpose	Availability
lex:	BM25 only	Exact keywords, identifiers, error strings	MCP + CLI
vec:	Vector only	Semantic rephrasing, concept search	MCP + CLI
hyde:	Vector only	Hypothetical document. Describe what the answer looks like	MCP + CLI
expand:	Full pipeline	Auto-expand into lex/vec/hyde sub-queries. Max one per query document.	CLI only
(plain text)	Full pipeline	Implicit expand:. Same as expand: but shorter to write.	CLI only

Fusion Weight

The first query line gets 2x fusion weight in RRF. Put your strongest signal first.

Examples

Single line, auto-expands (CLI only):

auth middleware

Multi-line typed query:

lex:handleAuth JWT token
vec:how does request authentication work
hyde:The middleware validates the JWT token from the Authorization header and attaches the decoded user object to the request context

Mixed, typed + auto-expand (CLI only, expand: not available via MCP):

lex:"ECONNREFUSED" retry logic
expand:how connection errors are retried

Lex Syntax

Lex queries support phrase matching and exclusions:

Syntax	Meaning	Example
"exact phrase"	Verbatim match	lex:"error boundary"
-term	Exclude term	lex:middleware -express
-"phrase"	Exclude phrase	lex:router -"page router"
bare words	OR match (stemmed)	lex:handleAuth parse token

Tools

Tool	Purpose
query	Search: typed sub-queries (lex, vec, hyde). CLI also supports expand: and plain text.
get	Retrieve one document by path
multi_get	Retrieve multiple documents
status	List collections, document counts, pending embeds

Context hygiene

Search results consume your main conversation context. For broad research across multiple collections, delegate to a subagent:

Use a subagent to research how Next.js handles routing
by searching the next.js QMD collection.

The subagent runs searches in its own context and returns a summary. Your main conversation stays clean.

For targeted lookups (one query, one collection), search directly. The overhead is low.

Retrieval

After search returns results, retrieve full documents:

Task	MCP Tool	Example
Get by path	get	file: "collection/path/to/doc.md"
Get by docid	get	file: "#abc123"
Get with line numbers	get	file: "docs/api.md", lineNumbers: true
Get from line offset	get	file: "docs/api.md", fromLine: 100, maxLines: 50
Multiple by glob	multi_get	pattern: "docs/*.md"
Multiple by list	multi_get	pattern: "#abc123, #def456"

MCP ↔ CLI Reference

MCP Tool	CLI Equivalent	Key Parameters
query	qmd query (alias: deep-search)	searches (array of {type, query} objects, types: lex/vec/hyde, min 1, max 10), limit (default 10), minScore (default 0), collections (array)
get	qmd get	file (path or #docid), fromLine, maxLines, lineNumbers
multi_get	qmd multi-get	pattern (glob or comma list), maxLines, maxBytes (default 10KB), lineNumbers
status	qmd status	(none)

CLI has additional search commands (qmd search, qmd vsearch) that map to BM25-only and vector-only searches. Use these in CLI fallback scenarios when you need a specific modality.

CLI defaults: 5 results in terminal mode, 20 for --json/--files. MCP defaults: 10.

CLI supports multiple collection filters: qmd search "query" -c react -c next.js. MCP collections parameter takes an array of collection names.

CLI output formats (not available via MCP): --files, --json, --csv, --md, --xml. MCP returns structured content via structuredContent field in addition to text.

Score Interpretation

Score	Meaning	Action
0.8 - 1.0	Highly relevant	Show to user
0.5 - 0.8	Moderately relevant	Include if few results
0.2 - 0.5	Somewhat relevant	Only if user wants more
0.0 - 0.2	Low relevance	Usually skip

Recommended Workflow

1. Check what's available: status
2. Search: query with lex/vec/hyde sub-queries
3. Refine: lex:"exact term" for identifiers, vec: for concepts, combine in one query document
4. Retrieve full documents: get with path or #docid

User asks how Next.js handles authentication middleware 1. query(searches: [{type: "lex", query: "authentication middleware"}, {type: "vec", query: "how auth middleware validates requests"}], collections: ["next.js"], limit: 10) 2. Review results, pick top 3 by score 3. get(file: "next.js/packages/next/src/server/web/adapter.ts", lineNumbers: true) 4. Summarize the pattern for the user User asks about a specific error in a known function 1. query with typed query document: lex:"NEXT_NOT_FOUND" handleNotFound vec:how the not-found error propagates through the router collections: ["next.js"], limit: 10 2. get for top result with lineNumbers User wants to compare error handling across React and Next.js 1. query(searches: [{type: "lex", query: "error boundary"}, {type: "vec", query: "error boundary implementation"}], collections: ["react"], limit: 5) 2. query(searches: [{type: "lex", query: "error handling middleware"}, {type: "vec", query: "how errors are handled in middleware"}], collections: ["next.js"], limit: 5) 3. get for top results from each 4. Compare approaches side-by-side User asks to research all routing patterns in Next.js Delegate to subagent (broad research consumes context): "Search the next.js QMD collection for routing patterns. Use query for 'routing', 'middleware', 'route handler', 'page router', 'app router'. Retrieve top results with get. Summarize findings."

Tips

• Run status first to see available collections and confirm zero pending embeddings.
• Scope searches with the collections array parameter when you know which repos to search.
• First query line gets 2x fusion weight. Put your strongest signal first.
• Use typed queries when auto-expand misses: lex: for exact identifiers, vec: for concepts, hyde: for "what would the answer look like".
• Max one expand: per query document (CLI only). Multiple lex:/vec:/hyde: lines are fine.
• After search returns a path, use @<absolute-path> to pull the full file into context (the collection path is in qmd status).
• If MCP tools are missing, the server may have disconnected. Run /qmd:status as a Bash fallback.
• Multi-get with glob: multi_get pattern "docs/*.md" or comma list "#abc, #def"
• Get with line offset: get file "docs/api.md:100" or use fromLine + maxLines params
• Context is hierarchical: all matching path prefixes are concatenated (global → root → specific), not just the deepest match
• For broad research, delegate to subagent to keep main context clean

CLI Fallback

When MCP tools are unavailable (server disconnected, not installed), use CLI:

qmd search "exact term" -c collection -n 10 --json     # BM25
qmd vsearch "concept" -c collection -n 10 --json       # Vector
qmd query "question" -c collection -n 10 --json         # Hybrid
qmd get path/to/file.md --line-numbers                  # Retrieve
qmd multi-get "docs/*.md" --json                         # Batch retrieve

CLI supports multiple -c flags: qmd search "query" -c react -c next.js

All 3 CLI search commands remain available. MCP has only query.

Named Indexes

QMD supports separate indexes via --index <name>. Config at ~/.config/qmd/<name>.yml, DB at ~/.cache/qmd/<name>.sqlite. Useful for work/personal separation:

qmd --index work search "deployment"
qmd --index personal search "journal"

Recovery

Problem	Cause	Fix
MCP tools missing from tool list	Server disconnected mid-session	Run /qmd:status (Bash fallback) or /mcp to check connection
Zero results after successful embed	Mask excluded key files	Re-add with /qmd:add <repo> --mask "<broader-glob>"
Search hangs or times out	Index corruption or model loading	Run qmd cleanup then qmd embed -f to rebuild
Low-quality results	Query too vague or wrong approach	Use typed queries: lex: for exact terms, vec: for concepts, combine both
"Collection not found" error	Typo or collection was removed	Run status to list available collections

Gotchas

• minScore: 0.5 filters too aggressively for broad queries. Use 0.3 for exploratory search.
• BM25 (lex) is better for exact identifiers. Vec is better for concepts. Combine both for best recall.
• expand: and plain text auto-expand are CLI-only. MCP requires explicit typed queries.
• Chunks are 900 tokens with 15% overlap. Search matches point to chunks, not exact lines. Use get with fromLine/maxLines to narrow down.