Graph pathfinding: connect action fails to find paths between related concepts

[aaronsb]

Observed Behavior

When using the MCP concept tool with action: connect to find paths between semantically related concepts, the query fails even when the concepts exist in the same document.

from_query: "graph visualization tool"
to_query: "workstation architecture"
threshold: 0.75
max_hops: 4

Result:

{
  "error": "No concepts found matching 'workstation architecture' at 75% similarity"
}

Expected Behavior

These concepts appear in the same ADR (ADR for Web Application Workstation Architecture). The document literally says:

"The web application started as a graph visualization tool... This ADR restructures the web application from a single-purpose visualization tool into a multi-function workstation."

Expected: Find a path, possibly direct or through intermediate concepts like "web application", "visualization", etc.

Hypothesis

1. Threshold too strict: 75% similarity might not match "workstation architecture" as a phrase even though the concept exists
2. Edge gap: The concepts exist but no explicit relationship edges connect them
3. Embedding mismatch: The query phrase embeddings don't align well with stored concept embeddings

Investigation Steps

• Check if both concepts exist via direct search at lower threshold
• Examine what relationships exist on both concepts
• Test with exact concept IDs (connection_mode: exact) instead of semantic
• Review edge creation during ingestion - are intra-document concept relationships being created?

Context

Discovered during knowledge graph introspection session - trying to trace the evolution from "visualization tool" to "workstation".