BuildContext

Traverses concept relationships via multi-hop graph queries.

Parameters

• conceptName (string, required): CONCEPT name (NOT a file). Example: "Machine Learning"
• depth (int, optional): Graph hops (default: 2). Range: 0+ (negative values error)
• maxEntities (int, optional): Max related entities to return (default: 20)
• includeContent (bool, optional): Include short content previews for source files (default: false)

Returns

Example

Constraints

• No hard requirement to Sync: If the graph is empty or the concept doesn't exist yet, the tool returns an empty result (no error). Run Sync to populate/update the graph.
• Depth semantics:
  • depth = 0 and depth = 1 return only directRelations.
  • depth > 1 also returns expandedRelations.
• ExpandedRelations traversal: Expansion starts from up to the top 5 direct relations (by co-occurrence), then traverses outward.
• includeContent behavior: Adds a ~200 character sentence-aware preview for up to 3 source files per direct relation; unreadable files are skipped silently.
• WikiLink normalization: [[Machine Learning]] → machine-learning

Decay Weighting

Concept relationships are weighted by recency:

• Each RelatedConcept includes DecayWeight (average across source files)
• WeightedScore = CoOccurrenceCount * DecayWeight
• Results are sorted by WeightedScore, not raw co-occurrence

This means recent concept connections rank higher than stale ones, even if the stale relationship has more historical co-occurrences.

Note: BuildContext does NOT update last_accessed. Use ReadMemory on specific files to boost their recency.

Community Detection

When community data is available (populated by Sync running the Louvain algorithm), BuildContext enriches its output in two ways:

CommunityId on RelatedConcepts

Each RelatedConcept in directRelations includes an optional communityId (int or null). This is the Louvain community cluster the concept belongs to. When no community data exists, the field is omitted.

CommunitySiblings

A separate communitySiblings array on the result surfaces concepts in the same community as the query concept that have no direct edge to it. These are structurally related concepts that edge traversal misses.

Each sibling includes:

Field	Type	Description
name	string	Concept name
communityId	int	Community cluster ID
sharedNeighborCount	int	Number of neighbors shared with the query concept
normalizedOverlap	double	sum(min(query_weight, sibling_weight) for shared neighbors) / sibling_total_degree

Siblings are filtered and capped by configurable thresholds:

Env Var	Default	Purpose
MAENIFOLD_COMMUNITY_MIN_SHARED	3	Minimum shared neighbors to qualify
MAENIFOLD_COMMUNITY_MIN_OVERLAP	0.4	Minimum normalized overlap score
MAENIFOLD_COMMUNITY_MAX_SIBLINGS	10	Maximum siblings returned

Graceful Degradation

When no community data exists (e.g., Sync hasn't run, or the graph is empty), communityId is null on all relations and communitySiblings is an empty array. No errors are raised. BuildContext falls back to its pre-community behavior.

Example with Community Data

Note: Community siblings are additive. Direct relations and expanded relations are returned as-is regardless of community membership. Siblings surface concepts that edge traversal misses — they don't filter or gate what edges return.

Integration

• Sync: Rebuild graph before BuildContext if files changed
• Visualize: Generate Mermaid diagrams from results
• SearchMemories: Find content after discovering concepts