Datasets:

JevlinAI
/

supersearch-tests

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
key: string
id: string
set: string
phrasing: int64
trial: int64
arm: string
query: string
request: struct<query: string, view: string, merged_limit: int64, max_results: int64, tier: string>
  child 0, query: string
  child 1, view: string
  child 2, merged_limit: int64
  child 3, max_results: int64
  child 4, tier: string
latency_ms: int64
first_attempt_ok: bool
ts_utc: timestamp[s]
response: struct<request_id: string, merged: list<item: struct<url: string, title: string, excerpt: string, fo (... 858 chars omitted)
  child 0, request_id: string
  child 1, merged: list<item: struct<url: string, title: string, excerpt: string, found_by: list<item: struct<provider: (... 63 chars omitted)
      child 0, item: struct<url: string, title: string, excerpt: string, found_by: list<item: struct<provider: string, ra (... 51 chars omitted)
          child 0, url: string
          child 1, title: string
          child 2, excerpt: string
          child 3, found_by: list<item: struct<provider: string, rank: int64>>
              child 0, item: struct<provider: string, rank: int64>
                  child 0, provider: string
                  child 1, rank: int64
          child 4, votes: int64
          child 5, published: timestamp[s]
  child 2, merged_total: int64
  child 3, decision: struct<tier: string, provider_set: string, providers: list<item: string>, build: struct<version: str (... 428 chars omitted)
      child 0, tier: string
      child 1, provider_set: string
      child 2, pr
...
     child 9, pool_pages: int64
          child 10, protected_pages: int64
          child 11, ids_dropped: int64
          child 12, exclusive_pages: int64
          child 13, top3_shared: struct<brave: double, exa: double, parallel-search: double, perplexity-search: double>
              child 0, brave: double
              child 1, exa: double
              child 2, parallel-search: double
              child 3, perplexity-search: double
      child 6, requested: null
  child 4, usage: struct<cost_micros: int64, latency_ms: int64, cache_hit: bool, deadline_hit: bool, byok: bool, rank_ (... 19 chars omitted)
      child 0, cost_micros: int64
      child 1, latency_ms: int64
      child 2, cache_hit: bool
      child 3, deadline_hit: bool
      child 4, byok: bool
      child 5, rank_cost_micros: int64
  child 5, by_provider: struct<>
  child 6, filters_applied: null
  child 7, no_usable_result: null
calls/balanced_pc600.jsonl.gz: string
phrasings.jsonl: string
README.md: string
calls/balanced_pc2000.jsonl.gz: string
results.json: string
calls/brave.jsonl.gz: string
examples.json: string
calls/parallel-search.jsonl.gz: string
chart.py: string
hf_export.py: string
calls/perplexity-search.jsonl.gz: string
build_items.py: string
examples.py: string
calls/fast.jsonl.gz: string
run_calls.py: string
calls/balanced.jsonl.gz: string
items.jsonl: string
report.py: string
calls/exa.jsonl.gz: string
calls/exa-lite.jsonl.gz: string
calls/balanced_pc4000.jsonl.gz: string
score.py: string
to
{'calls/balanced.jsonl.gz': Value('string'), 'calls/balanced_pc2000.jsonl.gz': Value('string'), 'calls/balanced_pc4000.jsonl.gz': Value('string'), 'calls/balanced_pc600.jsonl.gz': Value('string'), 'calls/brave.jsonl.gz': Value('string'), 'calls/exa-lite.jsonl.gz': Value('string'), 'calls/exa.jsonl.gz': Value('string'), 'calls/fast.jsonl.gz': Value('string'), 'calls/parallel-search.jsonl.gz': Value('string'), 'calls/perplexity-search.jsonl.gz': Value('string'), 'results.json': Value('string'), 'items.jsonl': Value('string'), 'phrasings.jsonl': Value('string'), 'examples.json': Value('string'), 'build_items.py': Value('string'), 'run_calls.py': Value('string'), 'score.py': Value('string'), 'chart.py': Value('string'), 'report.py': Value('string'), 'examples.py': Value('string'), 'hf_export.py': Value('string'), 'README.md': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              key: string
              id: string
              set: string
              phrasing: int64
              trial: int64
              arm: string
              query: string
              request: struct<query: string, view: string, merged_limit: int64, max_results: int64, tier: string>
                child 0, query: string
                child 1, view: string
                child 2, merged_limit: int64
                child 3, max_results: int64
                child 4, tier: string
              latency_ms: int64
              first_attempt_ok: bool
              ts_utc: timestamp[s]
              response: struct<request_id: string, merged: list<item: struct<url: string, title: string, excerpt: string, fo (... 858 chars omitted)
                child 0, request_id: string
                child 1, merged: list<item: struct<url: string, title: string, excerpt: string, found_by: list<item: struct<provider: (... 63 chars omitted)
                    child 0, item: struct<url: string, title: string, excerpt: string, found_by: list<item: struct<provider: string, ra (... 51 chars omitted)
                        child 0, url: string
                        child 1, title: string
                        child 2, excerpt: string
                        child 3, found_by: list<item: struct<provider: string, rank: int64>>
                            child 0, item: struct<provider: string, rank: int64>
                                child 0, provider: string
                                child 1, rank: int64
                        child 4, votes: int64
                        child 5, published: timestamp[s]
                child 2, merged_total: int64
                child 3, decision: struct<tier: string, provider_set: string, providers: list<item: string>, build: struct<version: str (... 428 chars omitted)
                    child 0, tier: string
                    child 1, provider_set: string
                    child 2, pr
              ...
                   child 9, pool_pages: int64
                        child 10, protected_pages: int64
                        child 11, ids_dropped: int64
                        child 12, exclusive_pages: int64
                        child 13, top3_shared: struct<brave: double, exa: double, parallel-search: double, perplexity-search: double>
                            child 0, brave: double
                            child 1, exa: double
                            child 2, parallel-search: double
                            child 3, perplexity-search: double
                    child 6, requested: null
                child 4, usage: struct<cost_micros: int64, latency_ms: int64, cache_hit: bool, deadline_hit: bool, byok: bool, rank_ (... 19 chars omitted)
                    child 0, cost_micros: int64
                    child 1, latency_ms: int64
                    child 2, cache_hit: bool
                    child 3, deadline_hit: bool
                    child 4, byok: bool
                    child 5, rank_cost_micros: int64
                child 5, by_provider: struct<>
                child 6, filters_applied: null
                child 7, no_usable_result: null
              calls/balanced_pc600.jsonl.gz: string
              phrasings.jsonl: string
              README.md: string
              calls/balanced_pc2000.jsonl.gz: string
              results.json: string
              calls/brave.jsonl.gz: string
              examples.json: string
              calls/parallel-search.jsonl.gz: string
              chart.py: string
              hf_export.py: string
              calls/perplexity-search.jsonl.gz: string
              build_items.py: string
              examples.py: string
              calls/fast.jsonl.gz: string
              run_calls.py: string
              calls/balanced.jsonl.gz: string
              items.jsonl: string
              report.py: string
              calls/exa.jsonl.gz: string
              calls/exa-lite.jsonl.gz: string
              calls/balanced_pc4000.jsonl.gz: string
              score.py: string
              to
              {'calls/balanced.jsonl.gz': Value('string'), 'calls/balanced_pc2000.jsonl.gz': Value('string'), 'calls/balanced_pc4000.jsonl.gz': Value('string'), 'calls/balanced_pc600.jsonl.gz': Value('string'), 'calls/brave.jsonl.gz': Value('string'), 'calls/exa-lite.jsonl.gz': Value('string'), 'calls/exa.jsonl.gz': Value('string'), 'calls/fast.jsonl.gz': Value('string'), 'calls/parallel-search.jsonl.gz': Value('string'), 'calls/perplexity-search.jsonl.gz': Value('string'), 'results.json': Value('string'), 'items.jsonl': Value('string'), 'phrasings.jsonl': Value('string'), 'examples.json': Value('string'), 'build_items.py': Value('string'), 'run_calls.py': Value('string'), 'score.py': Value('string'), 'chart.py': Value('string'), 'report.py': Value('string'), 'examples.py': Value('string'), 'hf_export.py': Value('string'), 'README.md': Value('string')}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Supersearch tests: one search index against a multi-index router

Raw data and scripts for a retrieval benchmark run on 27 October 2026. The question: when a coding agent searches the web, how often is the page that holds the answer in the top 10, for one search index on its own and for the Jevlin router's fan-out across several indexes. No model judge: the gold page is known for every item and matched by canonical URL.

Items (485)

obcode (30): OpenBenchmarks OB-Code-Websearch coding tickets; gold page from the # source: comment (CC-BY-4.0, openbenchmarks.com).
devdex (201): Firecrawl DevDex docs-track questions; gold = declared docs site + page stem or GitHub blob, DevDex's own rule (MIT).
seal (254): SealQA Seal-Hard questions; gold URLs with text fragments stripped (Apache-2.0).

items.jsonl has one row per item: id, set, query, gold_urls, and for DevDex gold_sites, gold_stems, gold_repo_files; gold_strings holds the answer text where the set provides one. phrasings.jsonl has three agent-style search queries per item, written by Claude Opus 5.5 from the question text only (words from the gold page slug removed).

Arms

Each provider was called on its own through the router (tier: balanced, providers: [x]): exa, exa-lite, brave, parallel-search, perplexity-search. fast and balanced are the router's live fused lists. balanced_pc600 and balanced_pc2000 are the balanced tier with page_chars 600 and 2000.

Calls (38315 successful records)

calls/<arm>.jsonl.gz (gzip JSON lines), one row per (item, phrasing, trial, arm): query, the exact request body, latency_ms of the successful attempt, first_attempt_ok, and response with the router's merged results (url, title, excerpt, found_by, votes), per-provider summary, usage (cost, cache hit, deadline) and the build hash. Trials 1 to 3 use max_results 10, 11, 12 so each is a fresh fan-out; trial 4 is a low-concurrency pass used for latency. Retries and failed attempts are not included.

Headline numbers (one query, exact URL, 485 items, trials 1 to 3)

arm	gold page in top 10	MRR@10	$ per search
exa	57.5	37.3	0.0069
exa-lite	49.7	34.5	0.004
brave	49.1	34.4	0.005
parallel-search	45.1	25.6	0.0039
perplexity-search	44.5	30.1	0.005
fast	60.3	38.9	0.0138
balanced	62.1	38.3	0.0204

Balanced minus exa, paired bootstrap over items: +4.6 [+1.5, +7.7]. Union of the four balanced arms (gold anywhere in their top-10 lists): 73.5%. Every pair and triple, the phrasing fan-out, the fast-tier curve, early-return policies and the page-length test are in results.json; concrete per-question examples are in examples.json.

Reproduce

python build_items.py (needs the three source datasets), python run_calls.py (needs a router key in JEVLIN_API_KEY), python score.py, python report.py results.json report.html. Canonicalisation and matching rules are documented at the top of score.py.

Downloads last month: 106

Total file size:

209 MB