NeuralWeb ARD · 19th Pinball
35 sources · 118,703 tables

ARD search across all 35 public sources

Which Open Knowledge Format (OKF) tables, from which of the 35 public sources (118,703 tables), can answer a question? Zero-LLM compiled reranking as a drop-in replacement for POST /search.

Sample questions

  • Corporate
  • Economics
  • US Census
  • Global Health
  • Disambiguation
  • Research Grants

Request · POST /search



        

Response


      

Benchmark: 197 Real NeuralWeb Queries

Measured across all 35 public ARD sources (118,703 OKF tables) on 197 real questions spanning Neural DB, Data Commons, and ARD.

Four-system comparison

Latency, cost per 1,000 searches, source retrieval accuracy, Foreign-NSI trap leakage, and independent blind evaluation.

System Latency Cost / 1k LLM Tokens Source Hit@5 Foreign-NSI Leak@5 Blind Judge
Loading benchmark metrics…

Source Hit@5 measures whether the expected source appears in the top 5 returned tables. Foreign-NSI Leak@5 measures single-country national statistics tables surfaced in the top 5 when that country was not mentioned in the question.

Explore all 197 real queries

Select any question to compare Flat Vector Search, Jev 1.14.0 LLM Reranker, and 19th Pinball (Zero-LLM Compiled) side by side.

197 queries

Select a query above

How 19th Pinball works & how it was trained

Built to serve as a drop-in, zero-LLM replacement for ARD search across all 35 public sources without sacrificing routing precision.

01 · OFFLINE TRAINING

Frontier listwise teacher across all 35 sources

Offline, a frontier model (gpt-6.2-sol) grades 2,995 diverse queries (59,900 candidate OKF pairs) side-by-side for entity, geography, grain, and metric match — strictly disjoint from all 197 locked evaluation queries.

02 · ONE UNIVERSAL MODEL

Shared-backbone reranker + compiled 19th .npz

Not a model per source: one universal model handles all 118,703 OKF tables. Zero-shot 35-source routing (Bumper 1) and hybrid dense + BM25 pooling (Bumper 2) feed a compiled 19th tabular .npz fused with either a fast 0.6B reader or a shared-backbone 4B LoRA adapter.

03 · LIVE SERVING

Zero LLM generation tokens at query time

Each POST /search embeds the question once, routes across 35 sources (~0.3 ms), pools candidates (~1–3 ms), and reranks with 0 LLM tokens and zero LLM fallback — drop-in compatible with ARD's JSON schema.