Query the llmwiki vault

Vault root: /mnt/c/Users/emmaj/llmwiki; CLI: llmwiki (on PATH; engines in .rag/, needs ollama serving qwen2.5:7b + nomic-embed-text).

Workflow

1. Locate — llmwiki search "<terms>": instant keyword scan, returns scored vault paths. Run this first; it tells you whether the vault covers the topic at all.
2. Read — open the top hit pages directly (they are small markdown files). For a factual question, reading 1–3 pages usually beats the generator and gives exact quotes plus frontmatter sources: for citation.
3. Ask — llmwiki ask "<question>": the AutoRAG-tuned pipeline (hybrid BM25+vector retrieval → local LLM). Use for synthesis questions spanning several pages, or when the user wants the pipeline's own answer. Takes ~10–60s.
4. Graph context — append --context to ask when the question is relational ("how does X connect to Y", entity inventories): it adds LightRAG knowledge-graph entities/relations. The graph covers only explicitly-graphed subsets (llmwiki graph <path> grows it); absence of graph hits is not absence of knowledge.
5. Answer honestly — the pipeline refusing ("context does not contain...") means the vault lacks the page, not that the answer doesn't exist. Say so and offer /ingest to close the gap.

Notes

• Cite vault pages by relative path (wiki/concepts/modding/Meatballing.md) so the user can open them in Obsidian.
• ask retrieves from trial-time indexes: pages added since the last tune are invisible to it (but visible to search). search finds it, ask refuses → invoke the retune skill; that mismatch is its trigger signature.
• The local 7B generator can smooth over details — for load-bearing claims, verify against the retrieved page text before asserting.