Vault Query Graph

vault-query-graph

Pure graph retrieval. Uses qkb graph neighbors and qkb graph query to traverse the wiki's typed link graph directly — no semantic retrieval, no rerank. Distinct from vault-query-qkb, which uses qkb query --graph (graph as one signal in a hybrid pipeline).

When this is the right tool:

• Entity-relation — "What does FAA NMS depend on?", "What feeds the Briefing Assembler?", "What links to NOTAM Reform?"
• Structural mapping — "Map the relationships around X" — you want the shape, not an answer.
• Walking from known seeds — you can name the starting entity/concept; the question is what's connected to it.

When this is NOT the right tool:

• Open-ended questions where you don't know which page to start from → use vault-query-qkb (the hybrid path will surface candidates).
• Single-page-answer questions ("what is X?") → use vault-query-bm25 or vault-query.
• Cross-source pattern questions where the answer isn't in any one page's wikilinks → use vault-query-qkb.

If the graph layer is empty (qkb graph status shows 0 nodes), fall back to vault-query-qkb.

Scope

• Vault root: /Users/dmestas/projects/<vault>-kb/wiki
• qkb index: default (~/.cache/qkb/index.sqlite)
• qkb collection: <vault> (pin -c <vault> for vault scope)
• Graph layer: ~1855 nodes / 14273 edges
• Edge types: LINKS_TO is the only edge type populated for this vault (this is normal for Obsidian vaults — Obsidian wikilinks are all LINKS_TO; EMBEDS and REFERENCES slots exist in qkb's schema but only fire for ![[file]] and frontmatter related:/sources: entries respectively, which are rarer)
• Node IDs come in two flavors:
  • doc:<N> — integer doc IDs assigned by the graph layer (not the same as the #<hash> FTS docids returned by qkb search)
  • wikitarget:<Page Name> — link targets that may or may not have a backing file yet (orphan wikilinks)
• Properties on nodes: id, path (where it exists), title. Label is :T (typed-node), not :Note.

Default workflow

Step 1 — Identify seed entities

Read the question. Pick 1–3 named entities or concepts to seed the graph walk. Examples:

• "How does Briefing Assembler relate to the NOTAM engine?" → seeds: Briefing Assembler, NOTAM (concept or entity)
• "What does FAA NMS depend on?" → seed: FAA NMS
• "Compare static vs transient data handling" → seeds: Transient Aeronautical Data (the conceptual cleavage page)

Step 2 — Resolve seed to the graph's node ID

Important: graph node IDs are NOT the same as FTS docids. qkb search returns hash-style ids (#4c5866); the graph layer uses integer-form doc:<N> ids. Use Cypher to find the graph node for a given path:

qkb graph query "MATCH (a) WHERE a.path = 'wiki/entities/FAA NMS.md' RETURN a.id LIMIT 1" --json
# returns e.g. doc:464

Or look up by title (less precise — may match multiple nodes with the same title across collections):

qkb graph query "MATCH (a) WHERE a.title = 'FAA NMS' RETURN a.id, a.path LIMIT 5" --json

Capture the integer-form doc:<N> id for each seed.

Step 3 — Walk the graph

For each seed, traverse outgoing edges:

qkb graph neighbors "doc:464" --hops 1 --edge-types LINKS_TO --json

Flags:

• --hops 1 — direct neighbors (start here). Bump to 2 only when the answer requires transitive connections.
• --edge-types — for Obsidian vaults LINKS_TO is usually all you need. Add EMBEDS,REFERENCES only if the vault makes heavy use of ![[file]] embeds or frontmatter related:/sources: arrays.
• --json — structured output.

For structural questions ("which concepts connect to X?"), prefer Cypher (more flexible than neighbors):

qkb graph query "MATCH (a)-[:LINKS_TO]->(b) WHERE a.path = 'wiki/entities/FAA NMS.md' RETURN b.id, b.title, b.path LIMIT 20" --json

For inbound edges ("what links TO X?"):

qkb graph query "MATCH (a)-[:LINKS_TO]->(b) WHERE b.path = 'wiki/entities/FAA NMS.md' RETURN a.id, a.title, a.path LIMIT 20" --json

For path between two entities:

qkb graph query "MATCH path = (a {path: 'wiki/entities/FAA NMS.md'})-[:LINKS_TO*1..3]->(b {path: 'wiki/concepts/Briefing Assembler.md'}) RETURN path LIMIT 5" --json

Step 4 — Filter and resolve to file paths

The neighbor list is the candidate set. Apply hygiene:

• Prefer: wiki/entities/, wiki/concepts/, wiki/sources/, wiki/comparisons/, wiki/questions/.
• Demote: wiki/hot.md, wiki/index.md, wiki/log.md, wiki/meta/* — these link broadly and add noise to a graph walk.
• De-dupe by resolved file path (graph and FTS can produce the same page via different routes).
• Cap at 5 reads even if the graph returns 20 neighbors — graph breadth doesn't justify unbounded reading.

Convert graph results to file paths:

qkb://<vault>/wiki/entities/foo.md
→ /Users/dmestas/projects/<vault>-kb/wiki/entities/foo.md

Watch for the hyphens-vs-spaces mismatch: graph results may show Schema-Drift-Manifest.md while the on-disk file is Schema-Drift Manifest.md. Use find wiki/ -iname '<name>.md' to resolve if a literal Read fails.

Step 5 — Read 3–5 selected pages, synthesize

Use the Read tool on the filtered candidates. For relational answers, the graph itself often carries the answer — describe what links to what, citing each surviving page:

[[Briefing Assembler]] consumes `RankedNotams` from the upstream grouper
(Source: [[Briefing Assembler]], LINKS_TO → [[Engine vs Assembler Boundary]]).

When citing, name the link type when relevant: "[[FAA NMS]] LINKS_TO [[NOTAM Reform]] and [[FAA SWIM]] — the dependency cluster."

Two-hop walks (use sparingly)

For "what does X transitively depend on?" questions, walk hops=2:

qkb graph neighbors "doc:<id>" --hops 2 --edge-types LINKS_TO --json

The candidate set explodes (10–50 nodes). Apply much tighter filtering — read the top-3 most-cited in the result (highest indegree from the seed), not the full breadth.

For genuinely deep traversal (transitive closure), use Cypher with path patterns — see the Step 3 examples above for the syntax (-[:LINKS_TO*1..3]->).

When the graph walk comes back empty or noisy

• Empty result from qkb graph neighbors — most likely you passed a hash-style FTS docid (#abc123) where the graph wants an integer-form id (doc:464). Use the Step 2 Cypher lookup to get the right id.
• Empty result with the right id — the seed page doesn't link out, or your seed lookup hit a wrong page. Try a different seed (a more central concept). Verify the page exists with find wiki/ -iname '<name>.md'.
• All meta-pages — the seed sits inside hot.md / index.md. Pick a more specific concept/entity as the seed instead.
• Genuinely sparse — the wiki doesn't link the concept densely yet. Fall back to vault-query-qkb which can use vector recall to find conceptual neighbors that aren't wikilinked.

If two retries don't surface useful neighbors, this is the wrong skill for the question. Stop and recommend vault-query-qkb.

Filing answers back

For non-trivial relational maps, offer to file as wiki/questions/<slug>.md:

---
type: question
title: "Short title"
question: "The exact question as asked."
answer_quality: solid
created: YYYY-MM-DD
tags: [question, graph, <domain>]
related:
  - "[[Page]]"
status: developing
---

After filing:

1. Add entry to wiki/index.md under Questions.
2. Append to wiki/log.md.
3. Run qkb update so the new page becomes graph-traversable.

Index maintenance

qkb update          # re-indexes + auto-runs graph link
qkb graph status    # node + edge counts

If the graph layer looks stale or wrong:

qkb graph link

(Runs automatically after every update; manual invocation rare.)

Guardrails

• Treat graph hits as candidates, not citations. Always Read the actual file before making source-backed claims. Edge existence ≠ semantic relevance.
• Don't use this skill for retrieval into the unknown. If you can't name a seed entity, you can't use the graph — that's vault-query-qkb's job.
• Resist transitive sprawl. Hops=2 is a strong claim that the answer is two wikilinks away. Most relational answers are one hop. Use hops=2 only when hops=1 came back insufficient.
• The graph layer is empty-corpus-safe — if qkb graph status shows 0 nodes, qkb graph neighbors returns nothing. Fall back to vault-query-qkb (the hybrid path's vector signal works regardless).
• Don't use qkb to ingest, rewrite, or save wiki pages. That's vault-save / vault-ingest territory.