What problem does it solve? Named entity recognition extracts entity mentions, but the same real-world person, place, or organization often appears under many spellings ("J. Smith", "John Smith", "Smith, J."), fragmenting relationship graphs and analysis. This Skill resolves that ambiguity by matching, scoring, and merging duplicate mentions into canonical entities with preserved aliases and a full audit trail. ## Core Features & Use Cases - Multi-Strategy Matching: Combines exact match, Jaccard token similarity, initial matching, Levenshtein edit distance, and phonetic (Soundex/Metaphone) matching with context boosters and reducers to compute composite confidence scores. - Conservative Merge Policy: Auto-merges only at 0.85+ confidence, routes 0.60-0.85 pairs to a human review queue, and never merges across entity types without explicit override. - Full Auditability and Reversibility: Preserves all aliases with provenance, logs every merge/split decision, and supports splitting entities when a merge is later proven wrong. - Use Case: After running NER over a new batch of legal documents, invoke resolve_entities to consolidate hundreds of name variants into canonical entities, then work through the review queue to confirm uncertain matches before building a relationship graph. ## Quick Start Ask the agent to run entity resolution on your document corpus after NER extraction, specifying the corpus ID and that uncertain matches should go to the review queue.
