weaviate/QueryAgentGym

Query Agent Gym

Query Agent Gym is a set of procedurally generated database environments and natural-language queries with exactly known ground-truth queries.

This release holds five held-out test environments containing 15 collections with 1,000 objects each (15,000 objects, 18,000 searchable documents) and three query sets over them. The default ask_mode set has 250 queries, 50 per environment. The search_mode set has 150 retrieval-only queries, 30 per environment, that test whether an agent picks a search, filters, or both. The edge_case_1 set has 130 cases that test exact-value filtering: whether an agent filters on an exact value or searches for it. Every entity, name, and document is synthetic.

Environment	Collections (kind)	Objects	Queries	Search mode	Edge case 1
Public Transit	TransitRoutes (knowledge_base), RidershipReports (transactional)	2,000	50	30	46
Tattoo Studios	TattooArtists, TattooDesigns (knowledge_base)	2,000	50	30	31
Nonprofits	Charities, Campaigns, GivingPrograms (knowledge_base), Donations (transactional)	4,000	50	30	23
Surf Schools	SurfBeaches, SurfInstructors, LessonPrograms, SurfEquipment (knowledge_base)	4,000	50	30	17
Childcare	DaycareCenters, ChildcareStaff (knowledge_base), ChildEnrollments (transactional)	3,000	50	30	13

Quick start

from datasets import load_dataset
import json

queries = load_dataset("weaviate/QueryAgentGym", "ask_mode", split="test")
surf = queries.filter(lambda r: r["environment"] == "surf-schools")
q = queries[0]
print(q["command"])                     # the natural-language request
print(q["ground_truth_sql"])            # readable ground truth, one entry per sub-query
plan = json.loads(q["ground_truth"])    # exact ground truth (query_database arguments)

routes = load_dataset("weaviate/QueryAgentGym", "public-transit_TransitRoutes", split="test")

search_mode = load_dataset("weaviate/QueryAgentGym", "search_mode", split="test")
print(search_mode[0]["retrieval_shape"])  # search_only | search_and_filters | filter_only

edge = load_dataset("weaviate/QueryAgentGym", "edge_case_1", split="test")
print(edge[0]["behavior"], edge[0]["command"], edge[0]["expect"])

Layout

Each environment has its own folder.

environments.parquet                   one row per environment
query_sets.parquet                     one row per query config (target mode, row format)
<environment>/schema.json              the schema exactly as generated
<environment>/queries.parquet          the environment's 50 ask mode queries
<environment>/search_mode_queries.parquet    the environment's 30 search mode queries
<environment>/edge_case_1.parquet            the environment's exact-value filtering cases
<environment>/objects/<Collection>.parquet   1,000 objects per collection

Config	Rows	Contents
ask_mode (default)	250	All five environments' queries, with an environment column
search_mode	150	All five environments' search mode queries, with environment and retrieval_shape columns
edge_case_1	130	All five environments' exact-value filtering cases, with behavior and expect columns
environments	5	One row per environment (its num_queries and queries_fingerprint describe the ask_mode set)
query_sets	3	One row per query config: the Query Agent mode it targets and how its rows are scored
<environment>_<Collection>	1,000 each	The objects of one collection, one typed column per property

To work with one environment, filter the ask_mode config on environment, or read <environment>/queries.parquet directly.

query_sets

An index of the query configs, so a benchmark loader can tell which Query Agent mode a set targets and how to score its rows without code changes.

Field	Type	Description
config	string	The query config, e.g. ask_mode
target_mode	string	The Query Agent mode the set targets: ask or search
record_format	string	ground_truth (rows carry the exact query_database arguments, as in ask_mode and search_mode) or expect (rows carry checks, as in edge_case_1)
description	string	One line on what the set tests
num_queries	int	Rows across the five environments
data_files	list of strings	The set's Parquet files

config	target_mode	record_format
ask_mode	ask	ground_truth
search_mode	search	ground_truth
edge_case_1	ask	expect

index = load_dataset("weaviate/QueryAgentGym", "query_sets", split="test")
for row in index:
    queries = load_dataset("weaviate/QueryAgentGym", row["config"], split="test")
    # run the agent in row["target_mode"], score per row["record_format"]

ask_mode

Field	Type	Description
query_id	string	e.g. public-transit-013
environment	string	Environment id, e.g. public-transit
environment_name	string	e.g. Public Transit
command	string	The natural-language request given to the agent
phrasing	string	explicit (names properties and values plainly) or conversational (a scenario, with values still verbatim)
plan	string	The operators of the plan, sub-queries separated by ;
num_sub_queries	int	1 to 3. Sub-queries are independent, never multi-hop
operator_classes	list of strings	Operator classes the plan features (search, each filter kind, each aggregation kind, group-by, multi-query)
ground_truth	string (JSON)	List of query_database argument objects, one per sub-query (format below)
ground_truth_sql	list of strings	A SQL-like rendering of each sub-query, for reading only
search_grounds	list of structs	For each sub-query with a search, its relevance-labeled result set (below)
reference_results	string (JSON)	Each sub-query's ground truth executed on the objects (below)

search_mode

The same fields as ask_mode, plus retrieval_shape. Every query is retrieval only, with no aggregations or group-by, and holds one sub-query. Each environment's 30 are split evenly across three retrieval shapes, and its query_ids run in shape order (000 to 009, 010 to 019, 020 to 029).

retrieval_shape	Per environment	The ground truth
search_only	10	A search and no filter
search_and_filters	10	A search and at least one filter
filter_only	10	At least one filter and no search

A search_only query's search is labeled over the whole collection, so its search_grounds[0].candidate_ids hold all 1,000 records.

edge_case_1

Each case checks one decision. The request gives an exact value, and the agent either filters on it or puts it in the search query. It should filter when the request names the property that holds the value, or when the value's form fits only one property. It should search when several properties could hold the value, or when the value may appear inside longer text. Rows carry checks rather than one exact answer.

Behavior	The request	Expected agent output	Node
B1	An exact value, naming the property that holds it	An = filter on that property. The search query is null when nothing else is asked	query
B2	An exact value that several properties could hold, or that appears inside longer text, without saying which	No filter on the value. The value appears verbatim in the search query	query
B3	An exact value, no property named, but only one property holds values of its form	An = filter on that property	query
B4	A prefix ("starts with")	LIKE '<text>*' on the property: the text, then a trailing *	query
B5	B1 or B2 plus another condition (category, date, number)	The other condition as an AND-ed filter, and the value handled as in B1 or B2	query
B6	A B1 lookup of a value that is not in the data	After zero results, keeps the = filter and does not fall back to search	retry
B7	A B2 request, after a filter on the wrong property returned nothing	Drops that filter and searches for the value verbatim	retry
B8	How many records involve a value that two properties can hold	An OR of one = filter per candidate property	aggregate
B9	A vague, descriptive term (a control)	A search query, or for a count at most one single-keyword LIKE. Never =	query, aggregate

Field	Type	Description
query_id	string	e.g. public-transit-b2-012
environment, environment_name	string	As in the other configs
behavior	string	B1 to B9
node	string	query, aggregate, or retry
collection	string	The collection the request targets
phrasing	string	quoted (the value in quotes), unquoted, or casual (a lowercase chat message, the value verbatim mid-sentence). B1 to B5 and B8 write every value in all three
command	string	The natural-language request
value	string	The exact value or prefix under test, verbatim from command. For B9 the topic word, matched case-insensitively
expect	struct	The checks (below)
relevant_ids	list of strings	The objects holding the value: in the named property (B1, B3, B5), in any candidate property or cited in free text (B2, B7, B8), or starting with the prefix (B4). B8's expected count is its length. Empty for B6 (the value is absent) and B9 (scored structurally only)
decoy_ids	list of strings	Objects holding a value one or two characters off (RID-00213 vs RID-00231), which a search may rank as matches. Usually empty for names
seed_output	string (JSON)	Retry cases only: the writer output to inject before the zero-results retry, {"collection", "search_query", "filters"}
source	string	Where the case comes from, e.g. same_form:routeName+comparisonRouteName, cited_in_text, named:reportCode

expect has these fields:

Field	Meaning
value_filter	How the value must be filtered: eq, like_prefix, or_eq, none (no = or LIKE on any property carries the value), or no_eq (B9 counts: no = on the term; a search or at most one single-keyword LIKE)
value_properties	The properties the value filter must target (for or_eq, all of them)
combine	OR for B8, otherwise null
query_contains	Strings that must appear verbatim in the search query
query_null	The search query must be null
other_filters	JSON list of filters that must also be AND-ed (B5), as query_database filter objects: {"property_name", "operator", "value"}
retry	keep (B6) or move_to_query (B7), else null

Where the cases come from. The data is used as published, and nothing was planted for this set:

• Public Transit's ridership reports name two routes per report (routeName, comparisonRouteName, same form), which gives the B2, B7 and B8 cases there.
• Tattoo designs cite artists in their descriptions, and staff and instructors are cited in each other's profiles, which gives B2 and B7 cases where the value only appears inside longer text.
• Donations cite their campaign's title in their notes as well as in campaignTitle.
• Codes (RID-, DON-, ENR-) and invented names give B1, B3, B4 and B6.

B8 and most B7 cases are in Public Transit, the only environment with two properties of one form. The cases follow an internal exact-value filtering spec, with these differences:

• All five environments are held out.
• There are 1,000 objects per collection rather than 2,000. That is still high cardinality, with 1,000 distinct identifiers and 424 to 440 distinct reference values.
• No decoys were planted, so most names have none.
• other_filters are filter objects rather than {property: value}, so date and number ranges can be expressed.
• Tokenization is left to the loader. Identifier properties should use Weaviate's field tokenization and free text word.

Objects

Each <environment>_<Collection> config has a uuid column followed by one column per property, in schema order. The uuid is the object's id and is what search_grounds and reference_results refer to. Property types map to columns and to Weaviate data types as follows.

Property type	Column type	Weaviate data type
searchable_text	string	text (the prose a search should match)
filterable_text	string	text (identifiers, references, categories)
int / number	int64 / float64	int / number
boolean	bool	boolean
date	string, YYYY-MM-DD	date (send as RFC 3339, e.g. 2025-11-16T00:00:00Z)
text_array, int_array, number_array, boolean_array, date_array	list	text[], int[], number[], boolean[], date[]

Each collection has exactly one identifier property (is_identifier in schema.json): an invented, unique name in a knowledge_base collection, a code in a transactional one. A transactional collection's reference properties (references in schema.json) hold the identifier of a knowledge_base record, stored directly in the row, so no query needs a join. No value is null.

Environment schemas

Every collection and property of the five environments, as in each schema.json. knowledge_base collections hold persistent, nameable things with an invented name as identifier, and transactional collections hold events that reference them. Values summarize what the objects hold: every value for a property with few distinct values (filters use these verbatim), otherwise the range, and for array properties how many items a record holds. A property's role is identifier (unique per record), reference to (another collection's identifier, stored in the row), or searchable (the prose a search should match).

Public Transit

A transit agency managing routes, vehicles, and ridership.

TransitRoutes (knowledge_base, 1,000 objects). Persistent transit routes operated by the agency.

Property	Type	Role	Description	Values
routeDescription	searchable_text	searchable	Describes the route's corridor, major stops, and communities served.	Prose, 118 to 168 words (mean 146)
routeName	filterable_text	identifier	The public-facing name of the transit route.	Unique per record, e.g. Morrowfen Link, Kestrelwick Crosstown
scheduledDepartureHours	int_array		The hours of the day, in 24-hour time, when scheduled departures occur.	5, 7, 12, 17, 23; 1 to 3 per record
stopSequenceNumbers	int_array		The ordinal positions of stops served along the route.	1, 4, 8, 13, 21; 1 to 3 per record

RidershipReports (transactional, 1,000 objects). Daily ridership reports for a transit route and a comparison route.

Property	Type	Role	Description	Values
reportCode	filterable_text	identifier	The unique code used to identify a daily ridership report.	Unique per record, e.g. RID-00001, RID-00002
routeName	filterable_text	reference to TransitRoutes	The public-facing name of the route whose boardings are reported.	A TransitRoutes.routeName value (432 distinct)
comparisonRouteName	filterable_text	reference to TransitRoutes	The public-facing name of the route used for comparison in the report.	A TransitRoutes.routeName value (424 distinct)
boardings	int		The total number of passenger boardings recorded for the reported route on the service date.	47 to 3100 (850 distinct)
serviceDate	date		The calendar date of the reported transit service.	2025-01-15 to 2026-09-18 (489 distinct)
vehicleAccessibilityChecks	boolean_array		Each flag indicates whether an observed vehicle accessibility check passed.	true or false (51% true; 1 to 2 per record)
fareEquipmentChecks	boolean_array		Each flag indicates whether an observed fare-equipment check passed.	true or false (50% true; 1 to 2 per record)
departurePunctualityChecks	boolean_array		Each flag indicates whether an observed departure met the agency's punctuality standard.	true or false (51% true; 1 to 2 per record)

Tattoo Studios

A tattoo studio network managing artists, designs, and appointments.

TattooArtists (knowledge_base, 1,000 objects). Artists working across the tattoo studio network, including their specialties and pricing.

Property	Type	Role	Description	Values
artistProfile	searchable_text	searchable	Describes the artist’s techniques, preferred subjects, and approach to collaborating with clients.	Prose, 124 to 170 words (mean 145)
artistName	filterable_text	identifier	The name clients use to identify and request the artist.	Unique per record, e.g. Maren Voss, Oren Pell
yearsExperience	int		The artist’s completed years of professional tattooing experience.	1 to 20 (20 distinct)
networkStartDate	date		The date the artist joined the studio network.	2018-04-19 to 2026-01-02 (835 distinct)
specialties	text_array		Tattoo styles and subjects the artist offers.	blackwork, botanical, fine line, traditional, watercolor; 1 to 3 per record
hourlyRates	number_array		Hourly rates in dollars for the artist’s available session offerings.	100, 125, 150, 185, 225; 1 to 3 per record
sessionDeposits	number_array		Deposit amounts in dollars accepted for the artist’s available session offerings.	50, 75, 100, 150, 200; 1 to 3 per record

TattooDesigns (knowledge_base, 1,000 objects). Tattoo concepts offered by network artists for clients to discuss and book.

Property	Type	Role	Description	Values
designDescription	searchable_text	searchable	Describes the design’s imagery, style, placement possibilities, and its association with an artist.	Prose, 126 to 176 words (mean 148)
designTitle	filterable_text	identifier	The title clients and artists use to identify the design.	Unique per record, e.g. The Lantern Moth, Saltglass Leviathan
placementSuitabilityFlags	boolean_array		Each flag indicates whether a placement considered for the design is suitable.	true or false (50% true; 1 to 2 per record)
sizeEstimatesCm	number_array		Estimated dimensions in centimeters for sizes in which the design can be tattooed.	5, 8.5, 12, 18, 25; 1 to 3 per record

Nonprofits

A nonprofit platform coordinating charities, campaigns, and donations.

Charities (knowledge_base, 1,000 objects). Nonprofit organizations that run fundraising campaigns on the platform.

Property	Type	Role	Description	Values
missionDescription	searchable_text	searchable	Describes the charity’s mission, programs, and communities served.	Prose, 122 to 174 words (mean 152)
charityName	filterable_text	identifier	The name by which supporters identify the charity.	Unique per record, e.g. Mosswake Watershed Circle, Hearthpan Community Table
causeCategory	filterable_text		The charity’s primary area of work.	Education, Environment, Food Security, Health, Housing
foundingYear	int		The year the charity was founded.	1978 to 2022 (45 distinct)
annualOperatingBudgetUsd	number		The charity’s annual operating budget in US dollars.	81,371.07 to 11,999,690.56 (1000 distinct)
suggestedDonationAmountsUsd	number_array		Donation amounts in US dollars that the charity suggests to supporters.	10, 25, 50, 100, 250; 1 to 3 per record

Campaigns (knowledge_base, 1,000 objects). Named fundraising initiatives that explain specific needs and ways to help.

Property	Type	Role	Description	Values
campaignStory	searchable_text	searchable	Explains the need the campaign addresses and the change it seeks to make.	Prose, 123 to 172 words (mean 148)
impactPlan	searchable_text	searchable	Describes how campaign funds are expected to produce measurable benefits.	Prose, 130 to 170 words (mean 150)
campaignTitle	filterable_text	identifier	The title supporters use to identify the campaign.	Unique per record, e.g. Moorbell School Pantry Lift, Gullhaven Senior Shoreline Renewal
campaignStatus	filterable_text		The campaign’s current fundraising status.	Active, Closed, Draft, Funded, Paused
acceptsRecurringGifts	boolean		Indicates whether supporters may set up recurring donations to the campaign.	true or false (50% true)
causeTags	text_array		Labels for the causes addressed by the campaign.	conservation, food access, housing stability, literacy, public health; 1 to 3 per record
beneficiaryTags	text_array		Labels for the people or communities the campaign aims to help.	coastal communities, families, rural residents, schoolchildren, seniors; 1 to 3 per record

GivingPrograms (knowledge_base, 1,000 objects). Named donation programs that define how supporters can give across the platform.

Property	Type	Role	Description	Values
programDescription	searchable_text	searchable	Explains the program’s giving approach and intended impact.	Prose, 124 to 168 words (mean 146)
programName	filterable_text	identifier	The name supporters use to identify the giving program.	Unique per record, e.g. Merevale Glasswing Ledger, Tallowmere Toolshare Circle
recurringEnrollmentAvailable	boolean		Indicates whether supporters can enroll in repeated giving through the program.	true or false (50% true)
employerMatchEligible	boolean		Indicates whether gifts through the program qualify for an employer matching process.	true or false (49% true)
launchDate	date		The date the giving program became available to supporters.	2021-03-16 to 2025-06-20 (729 distinct)
suggestedGiftAmountsUsd	number_array		Suggested gift amounts in US dollars for this giving program.	5, 15, 30, 75, 150; 1 to 3 per record

Donations (transactional, 1,000 objects). Individual donation transactions made to named fundraising campaigns.

Property	Type	Role	Description	Values
donorMessage	searchable_text	searchable	Records the donor’s optional message about the reason for giving.	Prose, 105 to 160 words (mean 137)
processingNotes	searchable_text	searchable	Records relevant context about processing or acknowledging the donation.	Prose, 95 to 147 words (mean 124)
donationCode	filterable_text	identifier	The short code a supporter or administrator uses to identify the donation.	Unique per record, e.g. DON-00001, DON-00002
campaignTitle	filterable_text	reference to Campaigns	The title of the campaign receiving the donation.	A Campaigns.campaignTitle value (430 distinct)
totalAmountUsd	number		The total amount charged for the donation in US dollars.	10.13 to 499.81 (987 distinct)
allocationAmountsUsd	number_array		The amounts in US dollars allocated to the donation’s recorded funding portions.	5, 10, 25, 50, 100; 1 to 3 per record

Surf Schools

A surf school managing beaches, instructors, and lesson bookings.

SurfBeaches (knowledge_base, 1,000 objects). Persistent beach locations used by the surf school for instruction.

Property	Type	Role	Description	Values
waveConditions	searchable_text	searchable	Describes the beach's typical wave behavior and suitable conditions.	Prose, 123 to 170 words (mean 152)
shoreAccess	searchable_text	searchable	Describes access to the water and facilities near the beach.	Prose, 130 to 172 words (mean 152)
safetyConsiderations	searchable_text	searchable	Describes recurring hazards and safety considerations at the beach.	Prose, 125 to 171 words (mean 147)
beachName	filterable_text	identifier	The beach's human-readable name.	Unique per record, e.g. Morrowglass Strand, Ternwick Cove
coastalArea	filterable_text		The coastal area containing the beach.	East Peninsula, Harbor Coast, North Shore, South Coast, West Bay
suitableAges	int_array		Ages for which the beach is generally suitable for supervised surf instruction.	8, 12, 16, 25, 40; 1 to 3 per record
typicalWaveHeightsMeters	number_array		Representative wave heights in meters under commonly encountered conditions.	0.3, 0.6, 0.9, 1.2, 1.8; 1 to 3 per record

SurfInstructors (knowledge_base, 1,000 objects). Instructors employed by the surf school to teach lessons.

Property	Type	Role	Description	Values
teachingApproach	searchable_text	searchable	Describes the instructor's teaching strengths and approach.	Prose, 127 to 170 words (mean 148)
instructorName	filterable_text	identifier	The instructor's human-readable name.	Unique per record, e.g. Merrith Vale, Orren Sable
hireDate	date		The date the instructor joined the surf school.	2017-03-15 to 2023-04-09 (812 distinct)
certificationExpiryDate	date		The expiry date of the instructor's current surf instruction certification.	2027-01-31 to 2029-02-28 (551 distinct)
teachingSpecialties	text_array		Types of surf instruction the instructor specializes in.	adaptive surfing, beginners, children, longboarding, wave reading; 1 to 3 per record
studentAgeGroups	int_array		Ages of students the instructor is qualified to teach.	8, 12, 16, 25, 40; 1 to 3 per record

LessonPrograms (knowledge_base, 1,000 objects). Persistent lesson offerings that define the surf school's instructional programs.

Property	Type	Role	Description	Values
curriculum	searchable_text	searchable	Describes the surfing skills and concepts taught in the program.	Prose, 124 to 169 words (mean 149)
equipmentGuidance	searchable_text	searchable	Describes the boards and other equipment suited to the program.	Prose, 124 to 175 words (mean 150)
programTitle	filterable_text	identifier	The lesson program's human-readable title.	Unique per record, e.g. Marshlight Ledger, Ternsilk Compass
durationMinutes	int		The length in minutes of a standard lesson in the program.	60 to 150 (91 distinct)
maximumStudents	int		The maximum number of students permitted in a lesson.	1, 2, 3, 4, 5, 6, 7, 8
packagePrices	number_array		Prices in dollars for the program's available lesson packages.	65, 95, 180, 275, 420; 1 to 3 per record
seasonAvailabilityDates	date_array		Published dates on which the program is offered during the current season.	2027-04-03, 2027-05-15, 2027-07-10, 2027-09-04; 1 to 3 per record

SurfEquipment (knowledge_base, 1,000 objects). Reusable surfboards and other named equipment assets maintained by the surf school.

Property	Type	Role	Description	Values
equipmentDescription	searchable_text	searchable	Describes the asset's construction, handling characteristics, and intended use.	Prose, 110 to 170 words (mean 143)
assetName	filterable_text	identifier	The equipment asset's human-readable name.	Unique per record, e.g. Morrowspindle, Ternwick Splay
equipmentCategory	filterable_text		The asset's equipment category.	Bodyboard, Fin Set, Leash, Surfboard, Wetsuit
replacementCost	number		The estimated replacement cost of the asset in dollars.	25.18 to 649.18 (994 distinct)
weightKilograms	number		The asset's weight in kilograms.	0.2 to 6.5 (489 distinct)
skillLevelTags	text_array		Surfer skill levels for which the asset is suitable.	advanced, beginner, first-timer, intermediate; 1 to 3 per record
materialTags	text_array		Materials used in the asset's construction.	epoxy, fiberglass, foam, neoprene, nylon; 1 to 3 per record
careTags	text_array		Handling and care requirements applicable to the asset.	fin inspection, freshwater rinse, shade storage, wax removal, zipper care; 1 to 3 per record
inspectionChecks	boolean_array		Each flag indicates whether an independently recorded equipment inspection check passed.	true or false (51% true; 1 to 2 per record)
plannedMaintenanceDates	date_array		Dates planned for preventive maintenance of the asset.	2027-03-01, 2027-05-01, 2027-07-01, 2027-09-01; 1 to 3 per record

Childcare

A childcare service coordinating daycare centers, staff, and enrollments.

DaycareCenters (knowledge_base, 1,000 objects). DaycareCenters holds the service's named childcare locations and their programs.

Property	Type	Role	Description	Values
programDescription	searchable_text	searchable	Describes a center's learning activities, care approach, and facilities.	Prose, 116 to 167 words (mean 144)
centerName	filterable_text	identifier	The name families and staff use to identify the center.	Unique per record, e.g. Mallowmere House, Tansybell House
acceptingEnrollments	boolean		Indicates whether the center is accepting new enrollment requests.	true or false (47% true)
providesMeals	boolean		Indicates whether the center provides meals during care.	true or false (51% true)
ageGroups	text_array		Lists the age groups served by the center.	infants, pre-K, preschoolers, toddlers; 1 to 3 per record
programFeatures	text_array		Lists features offered as part of the center's programs.	bilingual activities, music, outdoor play, sensory play; 1 to 3 per record
supportedLanguages	text_array		Lists languages supported in the center's daily communication.	English, French, Mandarin, Spanish; 1 to 3 per record
accessibilityFeatureChecks	boolean_array		Each flag indicates whether an assessed accessibility feature is available at the center.	true or false (50% true; 1 to 2 per record)
plannedClosureDates	date_array		Lists dates when the center plans to close for holidays or staff development.	2026-11-26, 2026-11-27, 2026-12-25, 2027-01-01; 1 to 3 per record

ChildcareStaff (knowledge_base, 1,000 objects). ChildcareStaff holds named caregivers and educators available to work across the service.

Property	Type	Role	Description	Values
careExperience	searchable_text	searchable	Describes a staff member's childcare experience and teaching strengths.	Prose, 120 to 171 words (mean 143)
staffName	filterable_text	identifier	The name used to identify the staff member.	Unique per record, e.g. Mira Dovren, Tavin Rell
yearsOfExperience	int		The number of completed years the staff member has worked in childcare.	1 to 22 (22 distinct)
qualifications	text_array		Lists the staff member's childcare qualifications and training.	CPR, early childhood diploma, first aid, food safety, special education training; 1 to 3 per record
openingShiftChecks	boolean_array		Each flag indicates whether the staff member is available for an assessed opening shift.	true or false (50% true; 1 to 2 per record)
closingShiftChecks	boolean_array		Each flag indicates whether the staff member is available for an assessed closing shift.	true or false (51% true; 1 to 2 per record)
substituteShiftChecks	boolean_array		Each flag indicates whether the staff member is willing to cover an assessed substitute shift.	true or false (49% true; 1 to 2 per record)
preferredWeeklyHours	int_array		Lists weekly hour totals the staff member is willing to work.	16, 24, 30, 35, 40; 1 to 3 per record

ChildEnrollments (transactional, 1,000 objects). ChildEnrollments records each child's enrollment at a daycare center.

Property	Type	Role	Description	Values
enrollmentCode	filterable_text	identifier	The short code used to identify and discuss an enrollment record.	Unique per record, e.g. ENR-00001, ENR-00002
centerName	filterable_text	reference to DaycareCenters	The name of the daycare center where the child is enrolled.	A DaycareCenters.centerName value (440 distinct)
childAgeMonths	int		The child's age in completed months when the enrollment begins.	6 to 60 (55 distinct)
startDate	date		The date the child's care at the center begins.	2026-10-12 to 2027-03-15 (155 distinct)

Ground truth format

A ground truth is a list of sub-queries, each a query_database function-call argument object.

[{"collection_name": "TransitRoutes",
  "search_query": "observatory and weather access",
  "numeric_array_property_filters": [
    {"property_name": "stopSequenceNumbers", "operator": "CONTAINS_ALL", "values": [1, 8]}]}]

Argument	Items	Operators or metrics
collection_name		The target collection
search_query	string	A relevance search over the collection's searchable text
text_property_filters	property_name, operator, value	=, LIKE (pattern uses %, e.g. Nerith%)
numeric_property_filters	property_name, operator, value	=, <, >, <=, >=
date_property_filters	property_name, operator, value (ISO date)	=, <, >, <=, >=
boolean_property_filters	property_name, operator, value	=, !=
array_property_filters (text arrays), numeric_array_property_filters, boolean_array_property_filters, date_array_property_filters	property_name, operator, values	CONTAINS_ANY, CONTAINS_ALL
numeric_property_aggregations	property_name, metric	COUNT, MIN, MAX, MEAN, MEDIAN, MODE, SUM
text_property_aggregations	property_name, metric, top_occurrences_limit	COUNT, COUNT_DISTINCT, TOP_OCCURRENCES
boolean_property_aggregations	property_name, metric	COUNT, TOTAL_TRUE, TOTAL_FALSE, PERCENTAGE_TRUE, PERCENTAGE_FALSE
date_property_aggregations	property_name, metric	COUNT, EARLIEST, LATEST
groupby_property	string	Groups every aggregation of the sub-query

All filters in a sub-query are combined with AND. A sub-query without aggregations retrieves records, and one with aggregations returns only the statistics.

Searches. The words of search_query are one valid phrasing and are not the target. A search sub-query's correct result is its labeled set: search_grounds[i].candidate_ids are the records the sub-query's filters select (R), and relevant_ids are the records a language model labeled relevant to the topic after reading every document in R (S). The topic and description fields say what the search is about.

Reference results. reference_results[i] is sub-query i executed on the objects, with a search resolving to its relevant_ids. A retrieval is a sorted list of uuids, an aggregation is {"METRIC(property)": value}, and a group-by maps each group value (as a string) to its aggregations. Percentages are on a 0 to 100 scale, and TOP_OCCURRENCES is a list of [value, count] pairs. Ties in MODE and TOP_OCCURRENCES are broken deterministically, which a database may break differently.

Evaluation

Placeholder, see github.com/weaviate/query-agent-benchmarking.

How the data was made

Each environment is a position in a seeded order over 200 domain briefs (seed 380, positions 5 to 9). Its structure is sampled procedurally: the number and kind of collections, property types, references, and each query's operator plan. Language models write the surface: the schema design (GPT-6 Sol), record names and documents (GPT-6 Luna), search topics and their relevance labels, and the natural-language commands (GPT-5.6 Luna). Ground truths are grounded in the saved objects and every command passes mechanical checks and a faithfulness verifier, so the plan can be recovered from the words. Each set of 50 queries is selected from a pool of 200 sampled plans so that every operator class appears at least 3 times, which makes the sets balanced across operators rather than distributed like natural traffic. A seeded coin chose each command's phrasing, giving 113 explicit and 137 conversational commands.

The search_mode queries were sampled as three separate pools of operator plans, one per retrieval shape, over search and the eight filter kinds only, with the same models, checks, and verifier. Each shape's 10 queries were selected from a pool of 200 sampled plans to cover the filter kinds the pool holds, and a seeded coin with even odds chose each command's phrasing.

The edge_case_1 cases were planned from the saved objects, with a fixed seed. Each value's relevant records were computed exactly, and each request was written by GPT-6 Luna under mechanical checks:

• the value appears verbatim in the requested style;
• the property is named only when the behavior calls for it;
• the records are referred to by their plain noun, with no schema vocabulary.

A request that kept failing its checks was replaced by the next candidate.

Limitations

• All data is synthetic and English only. Entities are invented, so an agent cannot answer from world knowledge, and documents are written as a real system of record would write them.
• Values are drawn independently, so a record can combine values that would rarely co-occur in reality (a fin set tagged for zipper care). Documents are written to stay consistent with each record's values.
• Public Transit has no group-by queries. Its only groupable properties are references with over 400 distinct values, which yield degenerate groups.
• Relevance labels for searches come from a language model reading each candidate document, not from human annotation.
• A search_only query's ground truth names only its collection and the fact that it searches, so an exact-match score cannot tell two search_only queries on the same collection apart. Their labeled result sets differ.
• A few search_only topics on the same collection cover overlapping themes (in Tattoo Studios, three architecture topics), so their relevant sets share many records.
• With 10 queries per shape, not every filter kind appears in every environment's search_mode queries.