onyx/backend/ee/danswer/db/analytics.py

import datetime
from collections.abc import Sequence
from uuid import UUID

from sqlalchemy import case
from sqlalchemy import cast
from sqlalchemy import Date
from sqlalchemy import func
from sqlalchemy import or_
from sqlalchemy import select
from sqlalchemy.orm import Session

from danswer.configs.constants import MessageType
from danswer.db.models import ChatMessage
from danswer.db.models import ChatMessageFeedback
from danswer.db.models import ChatSession


def fetch_query_analytics(
    start: datetime.datetime,
    end: datetime.datetime,
    db_session: Session,
) -> Sequence[tuple[int, int, int, datetime.date]]:
    stmt = (
        select(
            func.count(ChatMessage.id),
            func.sum(case((ChatMessageFeedback.is_positive, 1), else_=0)),
            func.sum(
                case(
                    (ChatMessageFeedback.is_positive == False, 1), else_=0  # noqa: E712
                )
            ),
            cast(ChatMessage.time_sent, Date),
        )
        .join(
            ChatMessageFeedback,
            ChatMessageFeedback.chat_message_id == ChatMessage.id,
            isouter=True,
        )
        .where(
            ChatMessage.time_sent >= start,
        )
        .where(
            ChatMessage.time_sent <= end,
        )
        .where(ChatMessage.message_type == MessageType.ASSISTANT)
        .group_by(cast(ChatMessage.time_sent, Date))
        .order_by(cast(ChatMessage.time_sent, Date))
    )

    return db_session.execute(stmt).all()  # type: ignore


def fetch_per_user_query_analytics(
    start: datetime.datetime,
    end: datetime.datetime,
    db_session: Session,
) -> Sequence[tuple[int, int, int, datetime.date, UUID]]:
    stmt = (
        select(
            func.count(ChatMessage.id),
            func.sum(case((ChatMessageFeedback.is_positive, 1), else_=0)),
            func.sum(
                case(
                    (ChatMessageFeedback.is_positive == False, 1), else_=0  # noqa: E712
                )
            ),
            cast(ChatMessage.time_sent, Date),
            ChatSession.user_id,
        )
        .join(ChatSession, ChatSession.id == ChatMessage.chat_session_id)
        .where(
            ChatMessage.time_sent >= start,
        )
        .where(
            ChatMessage.time_sent <= end,
        )
        .where(ChatMessage.message_type == MessageType.ASSISTANT)
        .group_by(cast(ChatMessage.time_sent, Date), ChatSession.user_id)
        .order_by(cast(ChatMessage.time_sent, Date), ChatSession.user_id)
    )

    return db_session.execute(stmt).all()  # type: ignore


def fetch_danswerbot_analytics(
    start: datetime.datetime,
    end: datetime.datetime,
    db_session: Session,
) -> Sequence[tuple[int, int, datetime.date]]:
    """Gets the:
    Date of each set of aggregated statistics
    Number of DanswerBot Queries (Chat Sessions)
    Number of instances of Negative feedback OR Needing additional help
        (only counting the last feedback)
    """
    # Get every chat session in the time range which is a Danswerbot flow
    # along with the first Assistant message which is the response to the user question.
    # Generally there should not be more than one AI message per chat session of this type
    subquery_first_ai_response = (
        db_session.query(
            ChatMessage.chat_session_id.label("chat_session_id"),
            func.min(ChatMessage.id).label("chat_message_id"),
        )
        .join(ChatSession, ChatSession.id == ChatMessage.chat_session_id)
        .where(
            ChatSession.time_created >= start,
            ChatSession.time_created <= end,
            ChatSession.danswerbot_flow.is_(True),
        )
        .where(
            ChatMessage.message_type == MessageType.ASSISTANT,
        )
        .group_by(ChatMessage.chat_session_id)
        .subquery()
    )

    # Get the chat message ids and most recent feedback for each of those chat messages,
    # not including the messages that have no feedback
    subquery_last_feedback = (
        db_session.query(
            ChatMessageFeedback.chat_message_id.label("chat_message_id"),
            func.max(ChatMessageFeedback.id).label("max_feedback_id"),
        )
        .group_by(ChatMessageFeedback.chat_message_id)
        .subquery()
    )

    results = (
        db_session.query(
            func.count(ChatSession.id).label("total_sessions"),
            # Need to explicitly specify this as False to handle the NULL case so the cases without
            # feedback aren't counted against Danswerbot
            func.sum(
                case(
                    (
                        or_(
                            ChatMessageFeedback.is_positive.is_(False),
                            ChatMessageFeedback.required_followup,
                        ),
                        1,
                    ),
                    else_=0,
                )
            ).label("negative_answer"),
            cast(ChatSession.time_created, Date).label("session_date"),
        )
        .join(
            subquery_first_ai_response,
            ChatSession.id == subquery_first_ai_response.c.chat_session_id,
        )
        # Combine the chat sessions with latest feedback to get the latest feedback for the first AI
        # message of the chat session where the chat session is Danswerbot type and within the time
        # range specified. Left/outer join used here to ensure that if no feedback, a null is used
        # for the feedback id
        .outerjoin(
            subquery_last_feedback,
            subquery_first_ai_response.c.chat_message_id
            == subquery_last_feedback.c.chat_message_id,
        )
        # Join the actual feedback table to get the feedback info for the sums
        # Outer join because the "last feedback" may be null
        .outerjoin(
            ChatMessageFeedback,
            ChatMessageFeedback.id == subquery_last_feedback.c.max_feedback_id,
        )
        .group_by(cast(ChatSession.time_created, Date))
        .order_by(cast(ChatSession.time_created, Date))
        .all()
    )

    return results