Write Path

Every write passes through auto-embedding, the admission gate, and conflict detection before being persisted.

Sequence

Step by step

1. Auto-embedding

If the write does not include a pre-computed vector and an Embedder is configured, the Content text is automatically embedded. The embedding is cached (LRU with SHA256 keys) to avoid redundant API calls. See Auto-Embedding for details.

2. Source resolution

Look up the source by ExternalID. If it doesn't exist, create one with neutral credibility (0.5). Apply label overrides ("moderator" -> 1.0, "troll" -> 0.05).

3. Near-duplicate scan

If the write includes a vector, do a quick ANN search for the 5 nearest existing nodes. This is used by the admission gate to detect duplicates and compute novelty.

4. Admission gate

Three rules run in order:

Rule	Condition	Result
Credibility floor	effective_credibility <= 0.05	Reject
Near-duplicate	max(similarity) >= 0.95	Reject
Novelty threshold	credibility * (1 - max_similarity) < threshold	Reject

5. Conflict detection

After admission, the conflict detector examines the nearest neighbours for contradictions. Candidates with moderate similarity (0.3–0.95) and shared labels are assessed. If an LLM provider is configured, it evaluates contradiction probability; otherwise, a heuristic is used. Confirmed contradictions create contradicts edges in the graph. See Conflict Detection for details.

6. Persist

If admitted:

• Graph store: UpsertNode writes the node with full metadata
• Vector index: Index adds the embedding for future ANN search
• Event log: Append records the write event
• Metrics: counters for admitted/rejected, latency histograms

7. Confidence assignment

The node's final confidence is:

node.confidence = initial_confidence * source_credibility

If no explicit confidence was provided, it defaults to the source's effective credibility.

8. Credibility feedback

The credibility learning background worker periodically reviews contradiction edges and adjusts source credibility via Bayesian updates. Sources that consistently produce validated information gain trust; sources that are frequently contradicted lose it.

IngestText pipeline

IngestText adds LLM extraction before the standard write path: