Every item ingested into Dump is automatically categorized by Gemini AI, which assigns a root category, subcategories, tags, summary, language, and key entities.

Root Categories

Dump uses 6 root categories:

Categorization Output

For each item, Gemini returns:

{
  category: string       // One of the 6 root categories
  subcategories: string[] // 0-3 specific subcategories
  tags: string[]         // 3-6 lowercase, hyphenated tags
  summary: string        // 1-2 sentence summary
  language: string       // ISO 639-1 code (e.g., "en", "pt")
  entities: Entity[]     // Up to 10 key entities
}

Entity Types

Each entity has a name and a type:

Model and Retry Logic

Dump tries models in order:

1. gemini-3.1-pro-preview
2. gemini-2.0-flash (fallback)

If the first attempt fails to parse valid JSON, it retries once. After two failures, the item receives the default category ("Reference") with empty subcategories and tags.

Embedding

After categorization, content is embedded using text-embedding-004:

• Model: Gemini text-embedding-004
• Dimensions: 768
• Max input: 2,000 characters (truncated)
• Failure mode: Non-fatal — item is saved without vector, full-text search still works