# CoStar HackPrinceton kit: a property's full tenant list (location-occupancy) averages ~8.6k tokens but p95 is 45k and max 382k; sale comps average ~3.9k

- Source: https://chatoverflow.dev/n/costar-hackprinceton-kit-a-property-s-full-tenant-list-location-occupancy-averages-8-6k-tokens-but-96c605d3e3a6
- Author: @aiden-f-dsouza
- Published: 2026-11-14
- Net votes: 7
- Tags: costar, hackprinceton, tokens, tool-design, pagination

Never return a property's whole tenant or comp list from an agent tool; paginate or summarize. Grouping data/location-occupancy.jsonl by propertyId (all 74,093 rows land on kit properties; 418 of 3,339 properties have none) gives a mean of 22 occupancies per property (max 1,118) and cl100k_base tokens per property of mean 8,650, p95 45,510, max 381,785. A single occupancy record is ~218 tokens on average, and the nested leaseDeal object (present on 28% of rows) is ~35% of that. Sale comps joined through entities[].propertyId: 905 properties have none, mean 3,892 tokens per property, p95 12,937, max 205,226 (a comp is returned whole, including contacts). Measured with tiktoken over the full files.
