LLM-Driven Extraction of Unstructured Data — Built for API Deployments & ETL Pipeline Workflows
Python ai-agentsdata-engineeringdocument-aigenerative-aiidpjson-extractionllmmcp-serverocrpdf-extractionprompt-engineeringstructured-output
"""Create, delete and share events for shareable resources.
Adapters, connectors, workflows, API deployments, pipelines and Prompt Studio
projects (plus any other model in ``SHAREABLE_RESOURCES``). The target type is
the resource's ``kind`` from that registry, so the read API and the share
endpoints name a resource type the same way.
"""
from dataclasses import dataclass
from typing import Any
from tenant_account_v2.shareable_resources import descriptor_for_kind, kind_for_instance
from platform_audit.recorder import best_effort, record_platform_event
# The field that says what sort of resource it is, where the kind alone does
# not: an LLM vs an embedding adapter, an ETL vs a Task pipeline.
_SUBTYPE_FIELDS = {
"adapter_instance": "adapter_type",
"pipeline": "pipeline_type",
}
@dataclass(frozen=True)
class ResourceRef:
kind: str
id: str
name: str
subtype: str
organization: Any
@best_effort
def resource_ref(resource: Any) -> ResourceRef | None:
"""What an event needs to name ``resource``, read now.
Read it before deleting: a deleted row has no primary key. Best effort,
because it runs before the delete: a failed read (``organization`` can be a
query) gives ``None``, which records nothing, rather than failing the delete.
"""
kind = kind_for_instance(resource)
if kind is None:
return None
descriptor = descriptor_for_kind(kind)
subtype_field = _SUBTYPE_FIELDS.get(kind)
return ResourceRef(
kind=kind,
id=str(getattr(resource, descriptor.id_field)),
name=str(getattr(resource, descriptor.name_field, "") or ""),
subtype=str(getattr(resource, subtype_field, "") or "") if subtype_field else "",
organization=getattr(resource, "organization", None),
)
@best_effort
def record_resource_event(
request: Any,
action: str,
resource: Any,
*,
summary: str = "",
changes: dict | None = None,
metadata: dict | None = None,
) -> None:
"""Record ``action`` on ``resource`` (a model instance or a ``ResourceRef``).
The summary defaults to the resource's subtype (``LLM``, ``ETL``).
"""
ref = resource if isinstance(resource, ResourceRef) else resource_ref(resource)
if ref is None:
return
record_platform_event(
action=action,
target_type=ref.kind,
target_id=ref.id,
target_name=ref.name,
summary=summary or ref.subtype,
changes=changes,
metadata=metadata,
request=request,
organization=ref.organization,
)
@best_effort
def record_sharing_change(
request: Any,
resource: Any,
*,
users: tuple[set, set] | None = None,
groups: tuple[set, set] | None = None,
shared_to_org: tuple[bool, bool] | None = None,
) -> None:
"""One ``resource.sharing_changed`` for a share request, if it changed anything.
Each argument is the ``(before, after)`` of an axis the request touched:
users and groups as model instances (stored as sorted e-mails and group
names), ``shared_to_org`` as booleans. Untouched axes are ``None``.
"""
from platform_audit.actions import Action
sides: dict[str, dict] = {"before": {}, "after": {}}
for axis, pair, label in (
("shared_users", users, lambda u: u.email),
("shared_groups", groups, lambda g: g.name),
):
if pair is None or pair[0] == pair[1]:
continue
sides["before"][axis] = sorted(label(x) for x in pair[0])
sides["after"][axis] = sorted(label(x) for x in pair[1])
if shared_to_org is not None and shared_to_org[0] != shared_to_org[1]:
sides["before"]["shared_to_org"] = shared_to_org[0]
sides["after"]["shared_to_org"] = shared_to_org[1]
if not sides["after"]:
return
record_resource_event(
request, Action.RESOURCE_SHARING_CHANGED, resource, changes=sides
)