Wiki Query

Steps

① Parse the question

Classify it — the strategy depends on type:

• Factual: "What is X?" → entity/concept lookup
• Comparative: "How does X differ from Y?" → check for existing comparison first
• Synthesis: "What do we know about X across sources?" → multi-page read
• Gap: "What don't we know about X?" → check open questions in pages, run wiki-lint

② Find relevant pages

Run wiki-search internally for key terms, inheriting the --tier default (active when tiers are enabled, else all). For wikis with 100+ pages, also filter by relevant tags from SCHEMA.md's taxonomy. Karpathy's rule: read index.md first, then drill into specific pages — this is exactly what wiki-search does.

If local search returns very few relevant pages, note that — step ⑤b (external fallback plugins) may apply. Don't skip the local read; local context still frames how the external results get synthesized.

③ Read the relevant pages

Use read_file. For large wikis, prioritize in order:

1. Existing comparison pages (if the question is comparative — may already answer it)
2. Concept pages (usually have the synthesis)
3. Entity pages (the specifics)
4. Query pages (previously filed answers that may contain the answer already)

④ Synthesize an answer

• Draw from compiled wiki knowledge, citing pages inline: "According to [[flash-attention]], the algorithm reduces memory from O(N²) to O(N)..."
• If the wiki has no coverage on part of the question, say so explicitly. Karpathy's "lint also finds content gaps" applies here — a gap in the answer is a signal for the next ingest.
• When pages contradict each other, present both views with their sources. Don't hide the contradiction.
• Match the structure to the question (prose / list / table / chart — see ⑤)

⑤ Apply output format

--format=markdown (default)

Prose with headers and [[wikilink]] citations throughout. The answer is a mini-article.

--format=table

Structured comparison table. Each row = option/entity, columns = dimensions:

| Dimension    | [[claude-3]] | [[gpt-4]]   | [[gemini]]  |
|--------------|--------------|-------------|-------------|
| Context      | 200k         | 128k        | 1M          |
| Multimodal   | yes          | yes         | yes         |
| Open weights | no           | no          | no          |
| Sources      | [[page-a]]   | [[page-b]]  | [[page-c]]  |

--format=slides (Marp)

Marp-compatible markdown slide deck. Obsidian has a Marp plugin; the Marp CLI also renders to PDF/HTML/PPTX.

---