Compose Corpus policies without hiding boundaries#
This example keeps the four policy dimensions separate while using
CorpusPolicyBundle as a reusable configuration
container. It performs no network request and loads no optional model.
Start from a safe local preset. The bundle is only composition: source execution, optional backends, downloader transport, and per-document failure behavior remain distinct policy objects.
policies = CorpusPolicyBundle.safe_local()
pprint(policies.explain())
assert policies.runtime.allow_network is False
assert policies.backend.allow_network is False
assert policies.backend.allow_download is False
assert policies.download.verify_ssl is True
assert policies.download.block_private_ips is True
{'backend': {'allow_download': False,
'allow_fallback': True,
'allow_network': False,
'include_unlisted': True,
'name': 'offline',
'on_exhausted': 'collect',
'order': [],
'require_ready': False},
'derived': {'backend_download_allowed': False,
'backend_fallback_allowed': True,
'backend_network_allowed': False,
'backend_requires_ready': False,
'pipeline_errors': 'collect',
'private_ip_blocking': True,
'tls_verification': True,
'url_sources_allowed': False},
'download': {'block_private_ips': True,
'max_bytes': 104857600,
'max_redirects': 5,
'max_retries': 3,
'name': 'secure',
'retry_backoff': 1.0,
'timeout': 30.0,
'user_agent': 'Mozilla/5.0 (compatible; scikitplot-corpus/1.0; '
'+https://github.com/scikit-plots/scikit-plots)',
'verify_ssl': True},
'errors': 'collect',
'name': 'safe-local',
'runtime': {'allow_network': False}}
Nested mapping configuration is useful for TOML/YAML/JSON frontends. Unknown keys and string booleans are rejected rather than silently treated as truthy.
team = CorpusPolicyBundle.from_config(
{
"preset": "safe-local",
"name": "team-local",
"backend": {
"preset": "offline",
"order": ["faster-whisper", "openai-whisper"],
"include_unlisted": False,
},
"download": {"preset": "constrained", "max_retries": 0},
"errors": "collect",
}
)
round_trip = CorpusPolicyBundle.from_config(team.to_dict())
assert round_trip == team
pprint(round_trip.to_dict())
{'backend': {'allow_download': False,
'allow_fallback': True,
'allow_network': False,
'include_unlisted': False,
'name': 'offline',
'on_exhausted': 'collect',
'order': ['faster-whisper', 'openai-whisper'],
'require_ready': False},
'download': {'block_private_ips': True,
'max_bytes': 26214400,
'max_redirects': 3,
'max_retries': 0,
'name': 'constrained',
'retry_backoff': 0.5,
'timeout': 15.0,
'user_agent': 'Mozilla/5.0 (compatible; scikitplot-corpus/1.0; '
'+https://github.com/scikit-plots/scikit-plots)',
'verify_ssl': True},
'errors': 'collect',
'name': 'team-local',
'runtime': {'allow_network': False}}
Helpers target only policy-owning seams. They do not push backend policy into every reader, because PDF/OCR/text/ASR fallback semantics are intentionally different.
reader_kwargs = team.reader_kwargs(model_size="tiny")
builder_config = team.builder_config(chunker="paragraph")
assert reader_kwargs["backend_policy"] == team.backend
assert builder_config.download_policy == team.download
assert builder_config.chunker == "paragraph"
One dimension can be tuned while the remaining contract stays immutable.
strict_team = team.with_overrides(name="team-strict", errors="raise")
assert strict_team.runtime == team.runtime
assert strict_team.backend == team.backend
assert strict_team.download == team.download
assert strict_team.errors.value == "raise"
print("Policy bundle example completed without network/model side effects.")
Policy bundle example completed without network/model side effects.
Total running time of the script: (0 minutes 0.007 seconds)
Related examples
Cleanprompt Moderate: Policies, Entities, and the Vault’s Lifecycle
Cleanprompt Packs and Formats: Records, Configs and Whole Folders
