Source code for agentc.auditor

import couchbase.exceptions
import datetime
import logging
import pathlib
import pydantic
import pydantic_settings
import textwrap
import typing

from .provider import Provider
from agentc_core.activity import DBAuditor
from agentc_core.activity import LocalAuditor
from agentc_core.analytics import Kind
from agentc_core.defaults import DEFAULT_ACTIVITY_FOLDER
from agentc_core.defaults import DEFAULT_CATALOG_FOLDER
from agentc_core.defaults import DEFAULT_LLM_ACTIVITY_NAME

logger = logging.getLogger(__name__)

# On audits, we need to export the "kind" associated with a log...
Kind = Kind


[docs] class Auditor(pydantic_settings.BaseSettings): """An auditor of various events (e.g., LLM completions) given a catalog.""" model_config = pydantic_settings.SettingsConfigDict(env_prefix="AGENT_CATALOG_") llm_model_name: typing.Optional[str] = None """ Name of the LLM model used to generate the chat messages to-be-audited. This field can be specified on instantiation or on :py:meth:`accept()`. A model specified in :py:meth:`accept()` overrides a model specified on instantiation. """ agent_name: typing.Optional[str] = None """ Name of the agent tied to this auditor. This field can be specified on instantiation or on :py:meth:`accept()`. Agent name specified in :py:meth:`accept()` overrides a model specified on instantiation. """ conn_string: typing.Optional[str] = None """ Couchbase connection string that points to the audit logs. This Couchbase instance refers to a CB instance that possesses the audit log collection. This collection is automatically generated on :command:`agentc index`, so this field is most likely the same instance as the CB instance possessing the catalog. If there exists no local audit log location (e.g., this is deployed in a standalone environment) OR if ``$AGENT_CATALOG_CATALOG`` is not explicitly set, we will perform all :py:meth:`accept` commands directly on the remote audit log collection. This field **must** be specified with :py:attr:`username`, :py:attr:`password`, and :py:attr:`bucket`. """ username: typing.Optional[pydantic.SecretStr] = None """ Username associated with the Couchbase instance possessing the audit logs. This field **must** be specified with :py:attr:`conn_string`, :py:attr:`password`, and :py:attr:`bucket`. """ password: typing.Optional[pydantic.SecretStr] = None """ Password associated with the Couchbase instance possessing the audit logs. This field **must** be specified with :py:attr:`conn_string`, :py:attr:`username`, and :py:attr:`bucket`. """ bucket: typing.Optional[str] = None """ The name of the Couchbase bucket possessing the audit logs. This field **must** be specified with :py:attr:`conn_string`, :py:attr:`username`, and :py:attr:`password`. """ certificate: typing.Optional[str] = None """ Path to the root certificate for TLS associated with the Couchbase cluster. This field **can** be specified with :py:attr:`conn_string`, :py:attr:`username`, and :py:attr:`password` for secure connection verification with Couchbase cluster. """ catalog: typing.Optional[pathlib.Path] = None """ Location of the catalog path. This field is used to search for the catalog version. If this field is not set, we will defer to the default behavior of :py:class:`agentc.Provider`. """ auditor_output: typing.Optional[pathlib.Path] = None """ Local audit log file to write to. If this field and ``$AGENT_CATALOG_CONN_STRING`` are not set, we will perform a best-effort search by walking upward from the current working directory until we find the :py:data:`agentc_core.defaults.DEFAULT_ACTIVITY_FOLDER` folder and subsequently generate an audit log here. Audit log files will reach a maximum of 128MB (by default) before they are rotated and compressed. """ _local_auditor: LocalAuditor = None _db_auditor: DBAuditor = None _audit: typing.Callable = None @pydantic.model_validator(mode="after") def _find_local_log(self) -> typing.Self: if self.auditor_output is None: working_path = pathlib.Path.cwd() logger.debug( 'Starting best effort search for the agent catalog folder. Searching for "%s".', DEFAULT_CATALOG_FOLDER, ) # Iteratively ascend our starting path until we find the catalog folder. while not (working_path / DEFAULT_CATALOG_FOLDER).exists(): if working_path.parent == working_path: # We have reached the root. We cannot find the catalog folder. return self working_path = working_path.parent (working_path / DEFAULT_ACTIVITY_FOLDER).mkdir(exist_ok=True) self.auditor_output = working_path / DEFAULT_ACTIVITY_FOLDER / DEFAULT_LLM_ACTIVITY_NAME return self @pydantic.model_validator(mode="after") def _find_remote_log(self) -> typing.Self: if self.conn_string is None: return self # Make sure we have {username, password, bucket}. if self.username is None: logger.warning("$AGENT_CATALOG_CONN_STRING is specified but $AGENT_CATALOG_USERNAME is missing.") return self if self.password is None: logger.warning("$AGENT_CATALOG_CONN_STRING is specified but $AGENT_CATALOG_PASSWORD is missing.") return self if self.bucket is None: logger.warning("$AGENT_CATALOG_CONN_STRING is specified but $AGENT_CATALOG_BUCKET is missing.") return self return self @pydantic.model_validator(mode="after") def _initialize_auditor(self) -> typing.Self: if self.auditor_output is None and self.conn_string is None: error_message = textwrap.dedent(""" Could not initialize a local or remote auditor! If this is a new project, please run the command `agentc index` before instantiating an auditor. If you are intending to use a remote-only auditor, please ensure that all of the relevant variables (i.e., conn_string, username, password, and bucket) are set. """) logger.error(error_message) raise ValueError(error_message) # TODO (GLENN): We should also add support for specifying a catalog version directly (with validation). # To grab the catalog version, we'll instantiate a provider. provider = Provider( conn_string=self.conn_string, username=self.username, password=self.password, bucket=self.bucket, catalog=self.catalog, ) # Finally, instantiate our auditors. if self.auditor_output is not None: self._local_auditor = LocalAuditor( output=self.auditor_output, catalog_version=provider.version, model_name=self.llm_model_name, agent_name=self.agent_name, ) if self.conn_string is not None: try: self._db_auditor = DBAuditor( conn_string=self.conn_string, username=self.username.get_secret_value(), password=self.password.get_secret_value(), certificate=self.certificate, model_name=self.llm_model_name, agent_name=self.agent_name, bucket=self.bucket, catalog_version=provider.version, ) except couchbase.exceptions.CouchbaseException: logger.warning("Could not connect to the Couchbase cluster. Skipping remote auditor.") self._db_auditor = None # If we have both a local and remote auditor, we'll use both. if self._local_auditor is not None and self._db_auditor is not None: logger.info("Using both a local auditor and a remote auditor.") def accept(*args, **kwargs): self._local_auditor.accept(*args, **kwargs) self._db_auditor.accept(*args, **kwargs) self._audit = accept elif self._local_auditor is not None: logger.info("Using a local auditor (a connection to a remote auditor could not be established).") self._audit = self._local_auditor.accept elif self._db_auditor is not None: logger.info("Using a remote auditor (a local auditor could not be instantiated).") self._audit = self._db_auditor.accept else: # We should never reach this point (this error is handled above). raise ValueError("Could not instantiate an auditor.") return self
[docs] def accept( self, kind: Kind, content: typing.Any, session: typing.AnyStr, grouping: typing.AnyStr = None, timestamp: datetime.datetime = None, model_name: str = None, agent_name: str = None, **kwargs, ) -> None: """ :param kind: Kind associated with the message. See analytics.log.Kind for all options here. :param content: The (JSON-serializable) message to record. This should be as close to the producer as possible. :param session: A unique string associated with the current session / conversation / thread. :param grouping: A unique string associated with one "generate" invocation across a group of messages. :param timestamp: The time associated with the message production. This must have time-zone information. :param model_name: LLM model used with this audit instance. This field can be specified on instantiation or on accept(). A model_name specified in accept() overrides a model specified on instantiation. :param agent_name: Agent used with this audit instance. This field can be specified on instantiation or on accept(). An agent_name specified in accept() overrides a model specified on instantiation. """ model_name = model_name if model_name is not None else self.llm_model_name agent_name = agent_name if agent_name is not None else self.agent_name self._audit( kind=kind, content=content, session=session, grouping=grouping, timestamp=timestamp, model_name=model_name, agent_name=agent_name, **kwargs, )
[docs] def move( self, node_name: typing.AnyStr, direction: typing.Literal["enter", "exit"], session: typing.AnyStr, content: typing.Any = None, timestamp: datetime.datetime = None, model_name: str = None, agent_name: str = None, **kwargs, ): """ :param node_name: The node / state this message applies to. :param direction: The direction of the move. This can be either "enter" or "exit". :param session: A unique string associated with the current session / conversation / thread. :param content: Any additional content that should be recorded with this message. :param timestamp: The time associated with the message production. This must have time-zone information. :param model_name: LLM model used with this audit instance. This field can be specified on instantiation or on accept(). A model_name specified in accept() overrides a model specified on instantiation. :param agent_name: Agent used with this audit instance. This field can be specified on instantiation or on accept(). An agent_name specified in accept() overrides a model specified on instantiation. """ model_name = model_name if model_name is not None else self.llm_model_name agent_name = agent_name if agent_name is not None else self.agent_name if direction == "enter": content = dict(to_node=node_name, extra=content) elif direction == "exit": content = dict(from_node=node_name, extra=content) else: raise ValueError('Direction must be either "enter" or "exit".') self._audit( kind=Kind.Transition, content=content, session=session, timestamp=timestamp, model_name=model_name, agent_name=agent_name, **kwargs, )