Documentation
¶
Overview ¶
Package graph holds Mesh's in-memory knowledge graph. The node/edge shape is lifted from dockyard's internal/knowledge graph and adapted for markdown: identity is the frontmatter id (node id "note:<id>"), never the file path, so a rename never rots an edge or an agent citation (spec section 3.6).
Index ¶
- Constants
- func Tokenize(s string) []string
- type Edge
- type Graph
- func (g *Graph) AddEdge(e Edge)
- func (g *Graph) AddNode(n *Node)
- func (g *Graph) CountByKind() map[string]int
- func (g *Graph) DetectCommunities(maxRounds int) int
- func (g *Graph) EdgeCount() int
- func (g *Graph) Neighbors(id string) []Edge
- func (g *Graph) NewRanker() *Ranker
- func (g *Graph) Node(id string) (*Node, bool)
- func (g *Graph) NodeCount() int
- func (g *Graph) Nodes() []*Node
- func (g *Graph) RecomputeDegrees()
- func (g *Graph) RefsTo(id string) []Edge
- type Node
- type Ranker
- type ScoredNode
Constants ¶
const ( ConfExtracted = "EXTRACTED" ConfInferred = "INFERRED" ConfAmbiguous = "AMBIGUOUS" )
Edge confidence levels (from the dockyard/graphify extraction model).
Variables ¶
This section is empty.
Functions ¶
Types ¶
type Graph ¶
type Graph struct {
// contains filtered or unexported fields
}
Graph is a concurrent-safe adjacency list. Reads take the RLock; the index builder is the only writer in M0.
func NewSized ¶
NewSized preallocates the maps for a graph expected to hold about n notes. A note contributes roughly one note node plus a handful of heading/tag nodes and edges, so the hints below avoid the repeated rehashing that dominates a from-scratch rebuild of a large vault. Hints are advisory; the graph still grows past them. n <= 0 builds with no hint (the small-vault default).
func (*Graph) AddEdge ¶
AddEdge inserts a unique (source, target, relation) edge and bumps the degree of whichever endpoints already exist.
func (*Graph) AddNode ¶
AddNode inserts a node, or enriches a previously bare reference node with real path/label data once its owning file is parsed.
func (*Graph) CountByKind ¶
func (*Graph) DetectCommunities ¶
DetectCommunities partitions the graph with the Louvain method (greedy modularity maximization) and writes a community id onto every node. Louvain beats the old label-propagation default on the quality that matters here: it does not collapse weakly-bridged clusters into one giant community, so the orientation tools and the web-graph coloring show meaningful groups. The result is deterministic (nodes are processed in sorted-id order, ties never move a node) and communities are renumbered 0..k-1 by their smallest member id, so the output shape is identical to the previous implementation; only the groupings improve. maxRounds caps the local-moving passes per level (default 20). Returns the number of communities.
func (*Graph) NewRanker ¶
NewRanker builds the inverted statistics over every note node's label+attrs.
func (*Graph) RecomputeDegrees ¶
func (g *Graph) RecomputeDegrees()
RecomputeDegrees sets every node's Degree from its adjacency lists, making the value independent of node/edge insertion order. AddEdge only bumps endpoints that already exist at insertion time, so a graph assembled by interleaving AddNode and AddEdge (BuildGraph: a references edge to a not-yet-added later note never counts that note's inbound degree) would otherwise disagree with one built nodes-first (LoadGraph). Call this once after the graph is fully assembled so both paths agree.
type Node ¶
type Node struct {
ID string // note:<frontmatter-id> | tag:<name> | note:<id>#<anchor>
Kind string // note|heading|tag|external|...
Label string
NoteID string // owning note's frontmatter id
NotePath string // denormalized for fast file open; refreshed on rename
Anchor string // heading slug for fetch-by-anchor
SourceLoc string // "L<line>" for the editor deep link
Community int
Attrs map[string]any
Degree int
}
type Ranker ¶
type Ranker struct {
// contains filtered or unexported fields
}
Ranker precomputes corpus statistics over the scorable (note) nodes once, so per-query scoring is O(queryTerms x candidates) instead of arachne's O(corpus)-per-call. Rebuild it whenever the graph changes.
type ScoredNode ¶
ScoredNode pairs a node with its BM25 relevance (higher is better).
Source Files