query

name	query
description	Use when the user asks a question that should be answered from the Cortex knowledge graph — "what does the vault say about X", "find the rule for Y", or /query <question>. Runs cited retrieval over the vault and answers grounded in the returned notes, always citing sources.

Answer from the Cortex knowledge graph (cited)

You are the AI layer over Cortex retrieval. The cortex toolkit does the deterministic, hybrid lexical+semantic retrieval; you synthesize a grounded, cited answer from what it returns. You never answer from memory when the graph has the answer.

Procedure

1. Resolve the CLI. Build the path to toolkit/dist/cli.js (run npm run build in toolkit/ first if dist/ is missing). Run all commands from the vault directory.

2. Retrieve. Run node <cli> query "<question>" --json and parse the JSON:

   {
     "question": "...",
     "anchors": ["..."],
     "hits": [
       { "path": "...", "id": "...", "title": "...", "type": "...", "score": 0.0, "excerpt": "...", "source": "...", "via": "anchor|link" }
     ],
     "sources": ["..."]
   }
   

   • hits are the relevant notes, best first. via: "anchor" means a direct match; via: "link" means it was reached through a wikilink (graph neighborhood).
   • sources are the original source documents behind the hits — the citation set.

3. Read before answering when needed. The excerpt is a snippet. If a hit is central to the answer, read the note at its path for the full content rather than relying on the excerpt alone.

4. Synthesize a grounded answer.

   • Answer only from the retrieved notes. If the hits don't cover the question, say so plainly and suggest atomizing the relevant source (or widening the question) — do not fill the gap from general knowledge.
   • Cite every claim. Reference the note title and/or its source. End with a short Sources list drawn from sources and the hit paths.
   • Prefer notes with higher score and via: "anchor"; use via: "link" hits as supporting context.

5. If retrieval is empty. Report that nothing in the graph matches, name the closest anchors tried, and offer next steps (/atomize <source> to add the missing knowledge, or rephrase).

Notes

• Read-only. query never writes; this skill only reads the vault and the retrieval output.
• Semantic search is optional. If the embedding store is absent, retrieval degrades to lexical TF-IDF (the toolkit prints a one-line hint to run cortex embed). Answers are still valid; cross-language/paraphrase recall is just lower.
• Citations are mandatory — an answer without a source is a bug. The whole point of Cortex is that every answer traces back to the graph.