# Every table in San Francisco

4,438 places to eat and drink, each scored for five different people, because asking a question of every business in a city now costs about the same as a cup of coffee.

Built from the SF Treasurer's business register, which records only a name and an address. Eighteen questions per place, $0.001 each, every answer carrying its sources.

The table starts the way the public record does: a name and a neighbourhood, and nothing that tells you what the place is like. Press Enrich and all eighteen questions go out across all 4,438 places at 431 a second, filling the table as the answers land.

Places
0
Answers
0
Cost
$0.00
p50 latency
—
p95 latency
—
Wall clock
—

431 req/sec · In flight 0 · Done 0 / 0 · 429s 0

Who are you with?Run the enrichment to rank and filter

Narrow it down

Showing 0 of 0

Fact columns price, walk to rail, outside, late, family run, loud. Checkable against the source.

Judgement columns walk in, linger, regulars, every “who is this for” score, underrated. The model's read of what it found.

#
Place
Fit
Price

## Method

  • Source: SF registered business locations, pulled 2026-10-25, filtered to food and drink with a public footprint.
  • 18 questions per place, one call each. A single grounded search answers all 18, so 18 questions cost what 1 does.
  • Total: 4,438 calls, $4.44. Served from recorded answers, paced to the measured service: 431 req/sec sustained, p50 550ms, p95 1.66s, hedged retry on slow calls, capacity 429s at the observed rate.
  • Scores are reported as percentile rank within this dataset, not raw model output.
  • Audit: 150 places re-read against their own citations. 82% of factual answers held up; 1% described a different business of the same name.
  • Walk to rail is not model output. It is computed from coordinates against BART and Muni Metro station locations, because the model guessed it at 66% precision.

## Schema

  • place (fact): name as filed with the city
  • neighborhood (fact): from the city register
  • fit (judgement): percentile for the selected occasion, or the best of the five when none is selected
  • price (fact): typical dinner for one, before drinks
  • walk_in (judgement): odds of a table at 8pm Friday with no reservation
  • linger (judgement): whether you can sit three hours without being moved along
  • regulars (judgement): whether a year of monthly visits would make you known
  • walk_to_rail (computed): minutes to nearest BART or Muni Metro
  • character (mixed): institution, family run, underrated, outside, late, loud, tourist trap, tech lunch
  • source (fact): the citation the answers were formed from

## Current view

  • Ranked for: A first date
  • Filters: none
  • Matching: 0 of 0

## Top 0 for a first date

    ## The point

    No vendor sells a “would this be cruel for breaking bad news” column. At a tenth of a cent per business you stop looking for a dataset and just ask the question. These eighteen took $4.44 across the whole city.

    ## It knows the city

    Best for your parents: House of Prime Rib, Harris', Original Joe's. Best alone with a book: Caffe Trieste. Worst for a hard conversation: Zeitgeist, The Knockout, El Rio, every one a loud dive bar.

    ## Honest about it

    Audited against their own citations, factual fields held up 82% of the time with 1% about the wrong business. Judgement columns are opinion, but every one links the source it was formed from. Most lead back to SF Eater, The Infatuation and the SF Standard.