glebis/qmd-search

qmd Search

Search a local markdown knowledge base semantically with qmd. Five modes — BM25 keywords, vector similarity, hybrid (expansion + rerank), literal native-script grep, and a fused find — all running on-device. The key advantage over Obsidian's built-in search: it matches meaning, finds notes that share no words with the query, and works across languages (e.g. a Russian query retrieves English notes).

When to use which mode

• hybrid (query) — default. A real question or fuzzy intent ("how do I stop overengineering"). Best quality; first run downloads reranker/expansion models (~one-time slow).
• vector (vsearch) — fast concept lookup ("notes about embodied computing").
• BM25 (search) — an exact keyword, name, or filename. Instant, no model.
• grep (-m grep) — literal fixed-string ripgrep over the .md files. The audit path for proper nouns, transliterations, exact phrases, Russian stems/inflections, and absence checks. Bypasses the index; matches only the exact script/spelling you type.

Bilingual / proper-name rule (do not skip)

This vault is bilingual (English/Russian). The embedding model is decent for concepts but weak for proper nouns / specific entities, and BM25 only matches the script you type. So:

Never conclude "it's not in the vault" after one English semantic query. For names, people, pets, places, foreign terms, or bilingual topics:

1. Search semantically first (query / vsearch).
2. Generate likely native-script spellings/stems and try them, e.g. Ziggy → Зигги/Зиги, dog/pet → собак, пёс, щенок, питомц, животн. Use stems (собак catches собака/собаку/собаки), not just the nominative.
3. Run a literal pass before concluding absence: qmd-search.sh -m grep -n 20 "Зигги".
4. Use literal hits to disambiguate close names (e.g. Зигги the pet vs. Зигмунд Freud).
5. If everything fails, say "I didn't find it with these queries: …" and list the terms tried — not "it's not in the vault." Raise -n to ~20 for absence checks.

Primary usage — the wrapper

Use the bundled wrapper; it suppresses qmd's stderr spinner, formats results as score path (parsing qmd's JSON, so commas in filenames are safe), and makes a best-effort refusal to run during an active qmd embed (which would return empty results — override with --force):

~/.claude/skills/qmd-search/scripts/qmd-search.sh [-m query|search|vsearch|grep|find] [-n N] [-c COLLECTION] [--snippet] [--min-score X] [--json] [--full] <query...>

Examples:

qmd-search.sh "what helps with anxiety"                 # hybrid (default)
qmd-search.sh -m vsearch -n 8 "behavioral health from photos"
qmd-search.sh -m search sensorium                       # BM25 keyword
qmd-search.sh -m grep -n 20 "Зигги"                     # literal native-spelling / absence check
qmd-search.sh -m find "Зигги собака"                    # fused: semantic + literal in one call
qmd-search.sh --snippet "agent orchestration"           # rows + matching snippets
qmd-search.sh --min-score 0.5 "quarterly planning"      # drop low-relevance hits
qmd-search.sh --json "agent orchestration"              # structured output for further processing

After getting hits, read the top files directly (they are normal vault paths) or fetch slices with qmd get "<path>:<line>" -l <N>.

Setup / indexing (only if qmd status shows the vault is not indexed)

qmd collection add ~/Brains/brain --name brain        # index the vault
qmd context add qmd://brain "short description of the vault"
qmd embed                                              # build vectors; re-run until status shows 0 pending
qmd cleanup                                            # compact the index

Refresh after large edits: qmd update && qmd embed. Check health any time with qmd status.

Operational rules (do not skip)

• One embed at a time, and never search while embedding — both cause empty/garbage results. The wrapper guards searches; for manual qmd calls, check qmd status first.
• If embedding never reaches 0 pending, check disk space (df -h) — a full disk fails writes silently. See references/cli-reference.md → "Operational gotchas".
• Vector scores are modest (~0.4–0.6); judge by ranking, not the absolute number.

MCP (native tools) vs. the CLI wrapper

qmd ships an MCP server (qmd mcp, stdio) exposing tools query, get, multi_get, status. If it's registered in the host (e.g. .mcp.json), prefer the native query tool for hybrid search — it returns structured results with no spinner/JSON-parsing/exit-code quirks. Register with:

{ "mcpServers": { "qmd": { "command": "qmd", "args": ["mcp"] } } }

Use the wrapper (scripts/qmd-search.sh) when you need what MCP doesn't cover: BM25-only (search), vector-only (vsearch), the literal/native-script grep pass, the fused find mode, --snippet, or --min-score. The bilingual/proper-name rule above applies to both paths.

Quality / evals

evals/fixture.example.json + scripts/run-evals.sh run qmd bench to score search quality (precision/recall/MRR per backend). Baseline and interpretation: evals/BASELINE.md. Re-run after changing the wrapper, the index, or the embedding model; a drop vs. baseline is a regression.

Reference

Full command surface, query grammar (lex:/vec:/hyde:), output formats, models, and recovery steps are in references/cli-reference.md.