Show & Tell — Phase 26.1 ConceptGraph: Hierarchical Concept Representation

[web3guru888]

ConceptGraph Architecture

The Problem: Unstructured Knowledge

Traditional AI systems store knowledge as flat key-value pairs or unstructured text. This makes it impossible to reason about relationships between concepts — is a penguin a bird? Does a bird have wings? Can a penguin fly? These questions require hierarchical, relational concept representation.

Solution: Semantic Network + Prototype Theory

ConceptGraph implements a hybrid approach combining Quillian's semantic networks (1967) with Rosch's prototype theory (1975) and Gärdenfors' conceptual spaces (2000).

Data Flow

Input Knowledge
    │
    ▼
ConceptNode ──────► ConceptEdge (IS_A/HAS_A/PART_OF/...)
    │                    │
    ▼                    ▼
Property            Adjacency
Inheritance         Index
    │                    │
    ▼                    ▼
Spreading ◄──────── Graph
Activation          Traversal
    │                    │
    ▼                    ▼
Similarity ◄──────── Subsumption
Metric               Check
    │                    │
    └────────┬───────────┘
             ▼
      Query Results

Key Algorithms

Property Inheritance — Depth-first IS_A traversal with override semantics:

async def propagate_properties(self, concept_id: str) -> FrozenSet[str]:
    own = self._nodes[concept_id].properties
    inherited = set()
    for parent_id in self._nodes[concept_id].parent_ids:
        parent_props = await self.propagate_properties(parent_id)
        inherited |= parent_props
    return frozenset(own | (inherited - own))  # own shadows inherited

Spreading Activation — Quillian-style with exponential decay:

async def spreading_activation(self, source_id, decay=0.8, threshold=0.1):
    activations = {source_id: 1.0}
    queue = deque([(source_id, 1.0)])
    while queue:
        node_id, activation = queue.popleft()
        for edge in self._adjacency[node_id]:
            new_act = activation * decay * edge.weight
            if new_act > threshold and new_act > activations.get(edge.target_id, 0):
                activations[edge.target_id] = new_act
                queue.append((edge.target_id, new_act))
    return activations

Wu-Palmer Similarity — Path-based measure using LCA:

depth_lca = depth(LCA(a, b))
similarity = (2 * depth_lca) / (depth(a) + depth(b))

Integration Map

WorldModel (13.1) ───► ConceptGraph (26.1) ◄─── SemanticParser (19.1)
                            │
                            ▼
                    OntologyManager (26.2)
                            │
                            ▼
                    CommonSenseEngine (26.4)

Prometheus Metrics

Metric	PromQL Example
concept_graph_nodes_total	concept_graph_nodes_total
concept_graph_query_seconds	histogram_quantile(0.95, concept_graph_query_seconds_bucket)
concept_graph_activation_spread_hops	avg(concept_graph_activation_spread_hops_sum / concept_graph_activation_spread_hops_count)

---

Issue: #578 | Planning: #577