Best LLMs for Topic Taxonomy Matching

Output schema

Every answer on this task is checked against this JSON Schema, whichever model wrote it. An answer that doesn't fit counts as a model failure, and the call is retried on another model.

{
  "$defs": {
    "TopicMatch": {
      "description": "One cluster label resolved against the shared Topic vocabulary.\n\nEither `topic_key` names an existing Topic, or the two `proposed_*` fields\ndescribe a genuinely novel theme. Never both, never neither.",
      "properties": {
        "client_topic_id": {
          "description": "The id of the cluster's draft topic being resolved",
          "title": "Client Topic Id",
          "type": "integer"
        },
        "proposed_description": {
          "anyOf": [
            {
              "maxLength": 500,
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Only when topic_key is null: what the new Topic covers and where its boundary lies. REQUIRED alongside proposed_name — the description is what the next run discriminates on (max 500 characters).",
          "title": "Proposed Description"
        },
        "proposed_name": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "Only when topic_key is null: a name for a NEW Topic. It must be reusable across subjects — no company, product, ticker or person names — because it joins a vocabulary every client matches against.",
          "title": "Proposed Name"
        },
        "topic_key": {
          "anyOf": [
            {
              "type": "string"
            },
            {
              "type": "null"
            }
          ],
          "default": null,
          "description": "The key of an existing Topic from the supplied vocabulary, when one covers this cluster. Null only when nothing in the vocabulary fits.",
          "title": "Topic Key"
        }
      },
      "required": [
        "client_topic_id"
      ],
      "title": "TopicMatch",
      "type": "object"
    }
  },
  "description": "One decision per draft supplied.\n\n⚠ Several drafts resolving to the same `topic_key` is the NORMAL and\nwanted case — that is how grouping emerges. Measured across 32 completed runs,\na run's clusters collapse onto a median of 14 distinct Topics (max 51 of 60).",
  "properties": {
    "assignments": {
      "description": "Exactly one entry per draft in the input",
      "items": {
        "$ref": "#/$defs/TopicMatch"
      },
      "title": "Assignments",
      "type": "array"
    }
  },
  "title": "TopicMatchingOutput",
  "type": "object"
}