dqtlib

The Data Quality Tool for Agentic BI. Tells you the what and surfaces the why.

Unifies your scattered data into one source of truth. Upgrades your existing models, dashboards, and queries into a causal semantic layer you didn't have to write. Picks up on trends and surfaces business insights, all wrapped in a quality harness that puts guardrails on the AI so the reports it generates stay on-spec.

pip install dqtlib

The import name is dqt:

from dqt import Check, Runner, MemoryStore

Full documentation and examples: https://github.com/antonbarr-data/dqt

Quality

All 64 detectors are benchmarked against labeled synthetic datasets. Benchmark scripts and raw results live in the GitHub repo and are run on every release:

Median F1: 1.00 | Detectors with F1 >= 0.8: 52/64 | Detectors with F1 >= 0.6: 58/64

Per-detector breakdown: examples/benchmarks/results_summary.md

Reproduce by cloning the repo and running:

python scripts/run_benchmark_suite.py --quick
python scripts/generate_benchmark_summary.py

Adapters

Six adapters ship in v1.0. Nightly CI runs against live credentials for the cloud warehouses.

Adapter	CI
PostgreSQL (local)	bundled — no credentials needed
ClickHouse
Snowflake
BigQuery
Databricks
Local (DuckDB)	bundled — no credentials needed

Slack alerts

Run a check suite and post results to Slack in three lines:

from dqt import Runner, MemoryStore
from dqt.notifications import SlackNotifier

suite = Runner(MemoryStore()).run_suite(checks, adapter)
SlackNotifier().send_suite_report(suite, title="nightly orders checks")
# set SLACK_WEBHOOK_URL env var — no bot token required

Control which checks appear in the message with level:

# only failures
SlackNotifier().send_suite_report(suite, level="fail")

# warnings + failures (default)
SlackNotifier().send_suite_report(suite, level="warn")

# every check including passes
SlackNotifier().send_suite_report(suite, level="all")

Full reference: docs/api/notifications.md

Detector documentation

64 statistical detectors across 10 groups — drift, outliers, time series, distribution, information theory, pattern, referential, schema, basic, and custom.

Every detector has a structured page at docs/algorithms/<group>/<slug>.md covering:

• What it computes and its parameters
• When it works well and when it fails (with concrete failure-mode table)
• Default-threshold calibration — empirical FPR across six canonical data shapes (Normal, Lognormal, Poisson, Beta, Pareto, Exponential)
• Recommended thresholds per data shape
• Canonical citation and runnable Python API example

Browse the full catalog: docs/algorithms/README.md