手繰る (taguru) —
how an LLM pulls knowledge out through Taguru MCP

First, pick from the list which context — which knowledge base of meaning — to search.

tools/call
  name: "list_contexts"
  arguments: {}

{
  "total": 1,
  "contexts": [{
    "name": "sake",
    "description": "青嶺酒造という架空の酒蔵の知識（沿革・杜氏・方針・醸造工程）",
    "stats": {
      "associations": 12,
      "top_concepts": [
        {"label": "青嶺酒造", "count": 6},
        {"label": "日本酒の醸造", "count": 5}
      ]
    }
  }]
}

A human-written description and machine-computed stats sit on the same row. From this list alone, the LLM can decide the sake context is the one to ask.

Resolve the cue 「青嶺酒造」, extracted from the question, to its canonical concept name on the graph.

{
  "context": "sake",
  "cue": "青嶺酒造"
}

[{
  "name": "青嶺酒造",
  "score": 1.0,
  "tier": "lexical",
  "kind": "exact",
  "gloss": "青嶺酒造。代表銘柄は青嶺。創業年は1907年。所在地は長野県。杜氏は高瀬。"
}, {
  "name": "青嶺",
  "score": 0.5,
  "kind": "containment",
  "gloss": "青嶺。青嶺酒造の代表銘柄は青嶺。"
}]

kind: "exact" means the cue was a registered spelling as-is. The gloss summarizes the concept's main facts, so a draft of the answer is nearly visible already — but nothing has been verified yet. The second candidate, 青嶺 (the brand), merely overlaps as a string — a containment candidate at a low 0.5, correctly left aside.

Survey which labels the concept 青嶺酒造 carries and how many facts each holds — a scouting pass before pulling the whole profile on a hunch.

{
  "context": "sake",
  "concept": "青嶺酒造"
}

{
  "concept": "青嶺酒造",
  "as_subject": [
    {"label":"代表銘柄","count":1},
    {"label":"創業年","count":1},
    {"label":"所在地","count":1},
    {"label":"杜氏","count":1},
    {"label":"行う","count":1},
    {"label":"醸造工程","count":1}
  ],
  "as_object": []
}

Six labels are visible — small enough that fetching everything will do (on a large hub concept, this is where you would narrow by label and then query). The presence of the 行う ("does") label pays off in the next activate as a negated fact.

Spread out from 青嶺酒造, collecting the direct facts and the knowledge chained beyond them, ordered by strength. The workhorse for gathering related knowledge.

{
  "context": "sake",
  "origins": ["青嶺酒造"]
}

[
 {"strength":0.5,"path":["青嶺酒造"],
  "association":{"label":"杜氏",
   "object":"高瀬","weight":1.0,
   "attributions":[{"source":
    "docs/aomine.md","paragraph":1,
    "section":"概要","locator":null}]}},
 {"strength":0.5,"path":["青嶺酒造"],
  "association":{"label":"行う",
   "object":"大量生産","weight":-1.0,
   "attributions":[{"source":
    "docs/aomine.md","paragraph":2,
    "section":"経営方針","locator":null}]}},
 … 代表銘柄 (flagship brand), 創業年 (founded), 所在地 (location) follow at strength 0.5
 {"strength":0.0417,"path":
  ["青嶺酒造","日本酒の醸造"],
  "association":{"label":"最初の工程",
   "object":"洗米", …}}
 … the brewing-process chain continues two hops out, weaker
]

Directly connected facts line up at strength 0.5, with a one-hop path of ["青嶺酒造"]. The brewing process — a separate chain of associations — mixes in with its strength decayed as the path stretches to two hops: the expansion naturally reaches "so what does the brewery do", while more distant information is weighted down. The negative weight: -1.0 is the negated fact "does not mass-produce" — the sign that keeps it from being read as an affirmative claim. attributions arrive with section already resolved, so which section states each fact — 概要 (overview) or 経営方針 (management policy) — is known at this point too.

Verify the grounds for "the master brewer is 高瀬" against the original text itself — not a summary, not an association.

{
  "context": "sake",
  "source": "docs/aomine.md",
  "paragraph": 1
}

{
  "text": "杜氏は高瀬が務めている。高瀬は先代から技術を受け継ぎ、伝統製法を守り続けている。",
  "source": "docs/aomine.md",
  "section": "概要",
  "locator": null
}

paragraph is simply the paragraph: 1 value that rode the activate attributions, passed back as-is. The association holds only the skeleton 杜氏 = 高瀬; the flesh — "inherited the craft from his predecessor" — was never skeletonized and stays in the original text. Which is exactly why it only surfaces on citation.

After resolving 浸漬 (steeping) with resolve, first check what facts the graph side holds.

{
  "context": "sake",
  "concept": "浸漬"
}

{
  "concept": "浸漬",
  "as_subject": [
    {"label":"次の工程","count":1}
  ],
  "as_object": [
    {"label":"次の工程","count":1},
    {"label":"工程","count":1}
  ]
}

There is order ("after steeping comes koji making") but no quantity label answering "how many minutes" — the skeleton of the procedure is written as associations, while details like conditions and amounts are, by design, never forced into triples and stay on the text side. Stop walking the graph here and switch lanes.