Conflict Detection

When new information contradicts existing knowledge, contextdb detects the conflict at write time and records it as a graph edge. This enables downstream retrieval to surface contradictions and lets credibility learning adjust source trust.

How it works

The conflict detector runs as part of the write path, after the admission gate accepts a candidate node.

Candidate filtering

Not every similar node is a potential contradiction. The detector applies these filters:

Criterion	Range	Rationale
Cosine similarity	0.30 – 0.95	Too low = unrelated; too high = near-duplicate (caught by admission gate)
Label overlap	At least one shared label	Nodes about different topics rarely contradict
Different source	Preferred, not required	Same-source contradictions are still valid

Contradiction assessment

For each candidate that passes filtering, the detector estimates P(contradiction):

With LLM provider: The candidate and existing node are sent to the LLM with a prompt asking whether they contradict. The response is parsed as a probability.

Without LLM (heuristic fallback): A simple heuristic based on similarity and confidence:

P(contradiction) = (1 - similarity) * min(candidate.confidence, existing.confidence)

Nodes with moderate similarity (saying similar things differently) and high confidence on both sides are more likely to be genuine contradictions.

Contradicts edges

When a contradiction is confirmed (P > 0.5), the detector creates a directed edge:

new_node --[contradicts]--> existing_node

The edge carries:

• Confidence: the probability estimate from assessment
• ValidFrom: the time of detection
• Provenance linking back to both sources

These edges are visible via graph walk and can be used by retrieval strategies to downweight contested claims.

Write result

The WriteResult includes ConflictIDs, a list of node IDs that the new write contradicts. Callers can use this to:

• Surface conflicts to end users
• Trigger review workflows
• Adjust their own confidence in the new information

Credibility feedback loop

Contradictions feed into credibility learning. When a new node from a trusted source contradicts an existing node from a less trusted source, the less trusted source's credibility decreases. Over time, sources that consistently produce contradicted information are penalised.

Configuration

Conflict detection is automatic when a vector is provided with the write. No additional configuration is required. The LLM provider (if configured via Options.LLMProvider) enables higher-quality contradiction assessment.