REFERENCE
Tool Surface
Nine tools. Three purposes. Orient yourself in the graph, find and explore what you need, then verify what you found.
One tool. Call it first. It returns the shape of the graph, the state of its metrics, and where health warnings cluster. The dashboard for the entire knowledge base.
Returns the graph summary — domains, node kind distribution, metrics dashboard with freshness annotations, top health warnings, and knowledge debt count. The single tool that tells you what the graph contains, what's healthy, and what needs attention.
Orient is the "where am I?" tool. Every conversation about billing knowledge should start here.
Four tools for locating and exploring. Search finds nodes. Node expands a single result. Domain surveys a territory. Subgraph reveals connections between a set of nodes.
Takes a natural-language query and returns ranked results. Each result shows its match tier — exact, taxonomy, or keyword — so you know how it was found. Supports filters by node kind, domain, and epistemic status.
Takes a node ID and returns the full node detail plus all adjacent nodes grouped by edge relation. Use depth: 2 or depth: 3 for multi-hop traversal — returns a subgraph instead of a flat neighborhood.
Takes a domain name and returns every node in it, grouped by kind. Shows internal edge density and cross-domain connections — which other domains this one talks to and through which edge types.
Takes a set of node IDs and returns the induced subgraph — all edges where both endpoints are in the set, plus cross-edges reaching outside. Also supports path finding between two nodes via BFS.
The typical flow: search to find starting points, node to expand what looks promising, subgraph to see how your findings relate.
Two tools for verification and audit. Evidence traces provenance for a specific claim. Health reports on the entire graph.
Takes a fact or thesis node ID and walks its evidence chain — incoming grounds edges down to data nodes. Returns the chain, a confidence verdict (strong / degraded / ungrounded), and the reason. Also supports bulk confidence assessment for multiple nodes at once.
Full health report — warnings grouped by type with counts and examples, stale metrics with impact analysis (which facts depend on them), and knowledge debt ranked by connectivity (most impactful ungrounded facts first).
Trust tools answer the meta-question: not "what does the graph say?" but "how much should I believe what the graph says?"
Two tools for watching patterns and keeping the graph current.
Takes a taxonomy category — like "subscription-lifecycle" or "payment-block" — and returns weekly ticket volume trends, peak weeks, overall direction, and which graph atoms receive evidence from that category. Without a category, returns an overview of all categories ranked by volume.
Re-reads all .graph/ JSON files into memory. Call this after recompiling the knowledge graph source. Returns the new node and edge counts with timing. No server restart needed.
The tools compose into recurring patterns.
orient → search (find entry points) → node (expand the best match) → evidence (verify key claims). The explore-then-verify loop.
domain (survey all nodes) → node (drill into specific groups) → subgraph (see cross-domain connections). The map-then-zoom loop.
health (find warnings) → evidence (trace specific debt items) → trends (check if ticket data supports the claims). The trust-but-verify loop.
reload (after graph recompile) → orient (check new shape) → health (see if warnings improved). The refresh-and-check loop.
The tools are the interface. The layers are the intelligence.