recalllisted
Install: claude install-skill shivamkpr/claude-skills
<!-- Public copy. Personal paths, names and ids are replaced with placeholders. -->
# Recall — RAG context retrieval (v3: qmd + Graphify)
Two retrieval backends. Pick by what the question needs; they compose.
- **qmd** — semantic passage search over the vault. Best for "what do we know / decide / say about X" — meaning-level lookup that returns the exact lines.
- **Graphify graph-RAG** — relationship and structure. Best for "how does X connect to Y", "what depends on X", "trace the flow", "what's in the same cluster". Two graphs:
- Vault: `$VAULT/graphify-out/graph.json` — projects, decisions, research.
- Claude setup: `$HOME/.claude/graphify-out/graph.json` — skills, agents, rules, memory (how the setup itself is wired).
## Path A — qmd (semantic, default for content questions)
1. `qmd query "<the actual question, in natural language>" --json -n 5` (always cap — default 20 burns ~10x tokens).
2. From the JSON (docid, path, score, snippet, context), pick the top 3–5.
3. Fetch only the answering section: `qmd get "#<docid>:<line-10>:40"`. Full `qmd get "<path>"` only if the section isn't enough.
4. Answer citing page names. Quote only what's needed.
5. Weak results? Retry once as keyword: `qmd search "<2-4 keywords>"`.
6. Read each hit by kind. A brief or living page is current truth. A progress record (audit, results, debrief, scan, ledger, fixed-bug write-up) is a dated fact about what happened: it answers "what was done / found / measured" and stays in default sear