- Improvements on audio processing to limit CPU and memory usage

- Removed Portkey from the equation, and defined explicit monitoring using Langchain native code
- Optimization of Business Event logging
This commit is contained in:
Josako
2024-10-02 14:11:46 +02:00
parent 883175b8f5
commit b700cfac64
13 changed files with 450 additions and 228 deletions

View File

@@ -30,6 +30,14 @@ class BusinessEvent:
self.environment = os.environ.get("FLASK_ENV", "development")
self.span_counter = 0
self.spans = []
self.llm_metrics = {
'total_tokens': 0,
'prompt_tokens': 0,
'completion_tokens': 0,
'total_time': 0,
'call_count': 0,
'interaction_type': None
}
def update_attribute(self, attribute: str, value: any):
if hasattr(self, attribute):
@@ -37,6 +45,22 @@ class BusinessEvent:
else:
raise AttributeError(f"'{self.__class__.__name__}' object has no attribute '{attribute}'")
def update_llm_metrics(self, metrics: dict):
self.llm_metrics['total_tokens'] += metrics['total_tokens']
self.llm_metrics['prompt_tokens'] += metrics['prompt_tokens']
self.llm_metrics['completion_tokens'] += metrics['completion_tokens']
self.llm_metrics['total_time'] += metrics['time_elapsed']
self.llm_metrics['call_count'] += 1
self.llm_metrics['interaction_type'] = metrics['interaction_type']
def reset_llm_metrics(self):
self.llm_metrics['total_tokens'] = 0
self.llm_metrics['prompt_tokens'] = 0
self.llm_metrics['completion_tokens'] = 0
self.llm_metrics['total_time'] = 0
self.llm_metrics['call_count'] = 0
self.llm_metrics['interaction_type'] = None
@contextmanager
def create_span(self, span_name: str):
# The create_span method is designed to be used as a context manager. We want to perform some actions when
@@ -61,6 +85,9 @@ class BusinessEvent:
try:
yield
finally:
if self.llm_metrics['call_count'] > 0:
self.log_final_metrics()
self.reset_llm_metrics()
self.log(f"Ending span {span_name}")
# Restore the previous span info
if self.spans:
@@ -82,7 +109,7 @@ class BusinessEvent:
'document_version_id': self.document_version_id,
'chat_session_id': self.chat_session_id,
'interaction_id': self.interaction_id,
'environment': self.environment
'environment': self.environment,
}
# log to Graylog
getattr(logger, level)(message, extra=log_data)
@@ -105,10 +132,108 @@ class BusinessEvent:
db.session.add(event_log)
db.session.commit()
def log_llm_metrics(self, metrics: dict, level: str = 'info'):
self.update_llm_metrics(metrics)
message = "LLM Metrics"
logger = logging.getLogger('business_events')
log_data = {
'event_type': self.event_type,
'tenant_id': self.tenant_id,
'trace_id': self.trace_id,
'span_id': self.span_id,
'span_name': self.span_name,
'parent_span_id': self.parent_span_id,
'document_version_id': self.document_version_id,
'chat_session_id': self.chat_session_id,
'interaction_id': self.interaction_id,
'environment': self.environment,
'llm_metrics_total_tokens': metrics['total_tokens'],
'llm_metrics_prompt_tokens': metrics['prompt_tokens'],
'llm_metrics_completion_tokens': metrics['completion_tokens'],
'llm_metrics_total_time': metrics['time_elapsed'],
'llm_interaction_type': metrics['interaction_type'],
}
# log to Graylog
getattr(logger, level)(message, extra=log_data)
# Log to database
event_log = BusinessEventLog(
timestamp=dt.now(tz=tz.utc),
event_type=self.event_type,
tenant_id=self.tenant_id,
trace_id=self.trace_id,
span_id=self.span_id,
span_name=self.span_name,
parent_span_id=self.parent_span_id,
document_version_id=self.document_version_id,
chat_session_id=self.chat_session_id,
interaction_id=self.interaction_id,
environment=self.environment,
llm_metrics_total_tokens=metrics['total_tokens'],
llm_metrics_prompt_tokens=metrics['prompt_tokens'],
llm_metrics_completion_tokens=metrics['completion_tokens'],
llm_metrics_total_time=metrics['time_elapsed'],
llm_interaction_type=metrics['interaction_type'],
message=message
)
db.session.add(event_log)
db.session.commit()
def log_final_metrics(self, level: str = 'info'):
logger = logging.getLogger('business_events')
message = "Final LLM Metrics"
log_data = {
'event_type': self.event_type,
'tenant_id': self.tenant_id,
'trace_id': self.trace_id,
'span_id': self.span_id,
'span_name': self.span_name,
'parent_span_id': self.parent_span_id,
'document_version_id': self.document_version_id,
'chat_session_id': self.chat_session_id,
'interaction_id': self.interaction_id,
'environment': self.environment,
'llm_metrics_total_tokens': self.llm_metrics['total_tokens'],
'llm_metrics_prompt_tokens': self.llm_metrics['prompt_tokens'],
'llm_metrics_completion_tokens': self.llm_metrics['completion_tokens'],
'llm_metrics_total_time': self.llm_metrics['total_time'],
'llm_metrics_call_count': self.llm_metrics['call_count'],
'llm_interaction_type': self.llm_metrics['interaction_type'],
}
# log to Graylog
getattr(logger, level)(message, extra=log_data)
# Log to database
event_log = BusinessEventLog(
timestamp=dt.now(tz=tz.utc),
event_type=self.event_type,
tenant_id=self.tenant_id,
trace_id=self.trace_id,
span_id=self.span_id,
span_name=self.span_name,
parent_span_id=self.parent_span_id,
document_version_id=self.document_version_id,
chat_session_id=self.chat_session_id,
interaction_id=self.interaction_id,
environment=self.environment,
llm_metrics_total_tokens=self.llm_metrics['total_tokens'],
llm_metrics_prompt_tokens=self.llm_metrics['prompt_tokens'],
llm_metrics_completion_tokens=self.llm_metrics['completion_tokens'],
llm_metrics_total_time=self.llm_metrics['total_time'],
llm_metrics_call_count=self.llm_metrics['call_count'],
llm_interaction_type=self.llm_metrics['interaction_type'],
message=message
)
db.session.add(event_log)
db.session.commit()
def __enter__(self):
self.log(f'Starting Trace for {self.event_type}')
return BusinessEventContext(self).__enter__()
def __exit__(self, exc_type, exc_val, exc_tb):
if self.llm_metrics['call_count'] > 0:
self.log_final_metrics()
self.reset_llm_metrics()
self.log(f'Ending Trace for {self.event_type}')
return BusinessEventContext(self).__exit__(exc_type, exc_val, exc_tb)