feat(ai-chat): Add code logic for AI-based data chat
Add the first working code logic both in terms of backend and frontend-related tasks. Add a detailled system message for improved results. Add several UI improvements for result display and user information. Add text input field for direct SQL code comparison. The implementation of the openAI backend had to be changed due to strict rate limits of azure OpenAI free tier and was replaced with a regular openai API key.
This commit is contained in:
150
app/app.py
150
app/app.py
@@ -1,5 +1,8 @@
|
||||
import json
|
||||
from typing import Any, Dict, Tuple
|
||||
|
||||
import pandas as pd
|
||||
from app_styles import header_style
|
||||
from dash import (
|
||||
Dash,
|
||||
Input,
|
||||
@@ -13,16 +16,40 @@ from dash import (
|
||||
no_update,
|
||||
)
|
||||
from dash.exceptions import PreventUpdate
|
||||
|
||||
from .app_styles import header_style
|
||||
from .data_chat import send_message
|
||||
from .sql_utils import execute_query, test_db_connection
|
||||
from data_chat import send_message
|
||||
from sql_utils import execute_query, test_db_connection
|
||||
|
||||
external_stylesheets = ["https://codepen.io/chriddyp/pen/bWLwgP.css"]
|
||||
|
||||
app = Dash(__name__, external_stylesheets=external_stylesheets)
|
||||
app.index_string = header_style
|
||||
|
||||
notification_md = """
|
||||
**Hinweise:**
|
||||
|
||||
Aufgrund des sparsamen pricing Tiers kann es einige Sekunden dauern, bis die
|
||||
Verbindung zur Datenbank hergestellt wird. Im Falle eines Fehlers gern ein-zwei mal erneut
|
||||
versuchen.
|
||||
|
||||
GPT-4o kann einige Fehler machen. Sollte dies passieren wird eine Fehlermeldung angezeigt.
|
||||
In diesem Fall lohnt es sich oft, die Anfrage leicht verändert erneut zu stellen und evtl
|
||||
zusätzliche Informationen zu geben.
|
||||
|
||||
Das Modell ist dazu aufgefordert, den Output stets auf 100 Zeilen zu begrenzen.
|
||||
|
||||
Alle Daten sind komplett zufällig generiert und haben keine Beziehung zu realen Personen.
|
||||
|
||||
**Beispielfragen**:
|
||||
- Wie viele Kunden haben wir in Hannover?
|
||||
- Zeige alle Kunden in Bremen.
|
||||
- Berechne den gesamten Stromverbrauch aller Kunden in Magdeburg.
|
||||
- Zeige alle Kunden, die zwischen 2021 und 2022 mindestens 200 Kubikmeter Gas verbraucht haben.
|
||||
- Wie viele Kunden haben zwischen 2021 und 2022 weniger Strom verbraucht als zwischen 2022
|
||||
und 2023?
|
||||
|
||||
Weitere Informationen zu den Daten, dem Code sowie zur Nutzung befinden sich in der README im
|
||||
[GiTea Repository](https://gitea.captain.particlephysics.de/quadfaselt/grid_application).
|
||||
"""
|
||||
|
||||
err_style = {
|
||||
"height": "0px",
|
||||
@@ -42,6 +69,46 @@ err_style = {
|
||||
}
|
||||
|
||||
|
||||
def render_table(df: pd.DataFrame) -> dash_table.DataTable:
|
||||
"""Create a Dash DataTable from a pandas DataFrame.
|
||||
|
||||
Parameters
|
||||
----------
|
||||
df : pd.DataFrame
|
||||
The input DataFrame to be rendered as a table.
|
||||
|
||||
Returns
|
||||
-------
|
||||
dash_table.DataTable
|
||||
A Dash DataTable component with styled layout and pagination.
|
||||
"""
|
||||
tab = dash_table.DataTable(
|
||||
id="table",
|
||||
columns=[{"name": i, "id": i} for i in df.columns],
|
||||
data=df.to_dict("records"),
|
||||
page_size=10,
|
||||
style_table={
|
||||
"overflowX": "auto",
|
||||
"margin": "auto",
|
||||
"width": "96%",
|
||||
"margin-top": "20px",
|
||||
},
|
||||
style_cell={
|
||||
"minWidth": "100px",
|
||||
"width": "150px",
|
||||
"maxWidth": "300px",
|
||||
"overflow": "hidden",
|
||||
"textOverflow": "ellipsis",
|
||||
},
|
||||
style_header={
|
||||
"backgroundColor": "lightblue",
|
||||
"fontWeight": "bold",
|
||||
"color": "black",
|
||||
},
|
||||
)
|
||||
return tab
|
||||
|
||||
|
||||
def get_layout() -> html.Div:
|
||||
"""Generate the layout for a Dash application.
|
||||
|
||||
@@ -53,6 +120,7 @@ def get_layout() -> html.Div:
|
||||
- A header with title and logo
|
||||
- A textarea for user input (database chat)
|
||||
- A submit button for the database chat
|
||||
- A Notification about the connection time and LLM performance
|
||||
- An error message area
|
||||
- A loading spinner and output area for database responses
|
||||
- A section for direct SQL queries, including a textarea and submit button
|
||||
@@ -78,6 +146,10 @@ def get_layout() -> html.Div:
|
||||
],
|
||||
className="header-container",
|
||||
), # Header
|
||||
html.Div(
|
||||
"Ganz ohne SQL-Kenntnisse Daten zu Zählerstandmessungen unserer Kunden abrufen!",
|
||||
style={"margin-left": "20px", "font-weight": "bold", "font-size": "20px"},
|
||||
),
|
||||
dcc.Store(
|
||||
id="tmp-value", data=start_value, storage_type="memory"
|
||||
), # Store previous prompt
|
||||
@@ -94,6 +166,17 @@ def get_layout() -> html.Div:
|
||||
disabled=False,
|
||||
style={"margin-left": "20px"},
|
||||
), # Submit button
|
||||
dcc.Markdown(
|
||||
notification_md,
|
||||
style={
|
||||
"margin-left": "20px",
|
||||
"margin-top": "20px",
|
||||
"margin-right": "10px",
|
||||
"background-color": "#C7E6F5",
|
||||
"border-radius": "5px",
|
||||
"padding": "10px",
|
||||
},
|
||||
),
|
||||
html.Div(
|
||||
[html.P("Bitte eine neue Anfrage eingeben.")], id="error", style=tmp_style
|
||||
), # Error message (only visible if input is not updated but submit button is clicked)
|
||||
@@ -102,7 +185,7 @@ def get_layout() -> html.Div:
|
||||
type="default",
|
||||
children=[
|
||||
html.Div(
|
||||
"Hier erscheint die Antwort der Datenbank.",
|
||||
"Hier erscheint die Antwort des KI-Modells.",
|
||||
id="text-output",
|
||||
style={
|
||||
"whiteSpace": "pre-line",
|
||||
@@ -180,18 +263,28 @@ def update_output(n_clicks: int, value: str, data: str) -> Tuple[Any, Any, Dict[
|
||||
Updated output text, new stored value, and error style.
|
||||
"""
|
||||
global err_style
|
||||
print(f"Value: {value}")
|
||||
print(f"Data: {data}")
|
||||
db_connected = test_db_connection()
|
||||
if n_clicks > 0 and value != data and db_connected:
|
||||
result = send_message(value)
|
||||
err_style["height"] = "0px"
|
||||
return result, value, err_style, html.P("")
|
||||
elif value == data:
|
||||
|
||||
err_style["height"] = "50px"
|
||||
err_child = html.P("Bitte eine neue Anfrage eingeben.")
|
||||
return no_update, no_update, err_style, err_child
|
||||
# parse LLM response to dict, then try to execute the query
|
||||
try:
|
||||
parsed_result = json.loads(result, strict=False)
|
||||
|
||||
result_table = execute_query(parsed_result["query"])
|
||||
children = [
|
||||
html.P([html.B("Zusammenfassung: "), f"{parsed_result['summary']}"]),
|
||||
html.P([html.B("SQL Abfrage: "), f"{parsed_result['query']}"]),
|
||||
render_table(result_table),
|
||||
]
|
||||
return children, value, err_style, html.P("")
|
||||
except Exception as e:
|
||||
err_style["height"] = "400px"
|
||||
err_child = html.Div(f"Folgender Fehler ist aufgetreten: {e}.LLM Output: {result}.")
|
||||
|
||||
return no_update, value, err_style, err_child
|
||||
|
||||
elif not db_connected:
|
||||
err_style["height"] = "50px"
|
||||
err_child = html.P(
|
||||
@@ -202,6 +295,12 @@ def update_output(n_clicks: int, value: str, data: str) -> Tuple[Any, Any, Dict[
|
||||
)
|
||||
return no_update, no_update, err_style, err_child
|
||||
|
||||
elif value == data:
|
||||
|
||||
err_style["height"] = "50px"
|
||||
err_child = html.P("Bitte eine neue Anfrage eingeben.")
|
||||
return no_update, no_update, err_style, err_child
|
||||
|
||||
raise PreventUpdate
|
||||
|
||||
|
||||
@@ -231,37 +330,14 @@ def run_sql_query(n_clicks: int, value: str) -> str:
|
||||
if isinstance(result, str):
|
||||
global err_style
|
||||
tmp_style = err_style.copy()
|
||||
tmp_style["height"] = "80px"
|
||||
tmp_style["height"] = "100px"
|
||||
tmp_style["padding"] = "20px"
|
||||
err_child = html.Div(
|
||||
[html.P(f"Fehler bei der Ausführung der Abfrage: {result}")], style=tmp_style
|
||||
)
|
||||
return err_child
|
||||
else:
|
||||
table_child = dash_table.DataTable(
|
||||
id="table",
|
||||
columns=[{"name": i, "id": i} for i in result.columns],
|
||||
data=result.to_dict("records"),
|
||||
page_size=10,
|
||||
style_table={
|
||||
"overflowX": "auto",
|
||||
"margin": "auto",
|
||||
"width": "96%",
|
||||
},
|
||||
style_cell={
|
||||
"minWidth": "100px",
|
||||
"width": "150px",
|
||||
"maxWidth": "300px",
|
||||
"overflow": "hidden",
|
||||
"textOverflow": "ellipsis",
|
||||
},
|
||||
style_header={
|
||||
"backgroundColor": "lightblue",
|
||||
"fontWeight": "bold",
|
||||
"color": "black",
|
||||
},
|
||||
)
|
||||
return table_child
|
||||
return render_table(result)
|
||||
raise PreventUpdate
|
||||
|
||||
|
||||
|
||||
@@ -1,16 +1,24 @@
|
||||
import os
|
||||
|
||||
from openai import AzureOpenAI
|
||||
from openai import OpenAI
|
||||
|
||||
# from openai import AzureOpenAI
|
||||
|
||||
# Set up credentials
|
||||
# NOTE: When running locally, these have to be set in the environment
|
||||
client = AzureOpenAI(
|
||||
azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT"),
|
||||
api_key=os.getenv("AZURE_OPENAI_KEY"),
|
||||
api_version="2024-02-01",
|
||||
)
|
||||
# NOTE: Usually I would use AzureOpenAI, but due to heavy rate
|
||||
# limitations on azure trial accounts, I am using OpenAI directly
|
||||
# for this project. However, this is how it would look like for
|
||||
# AzureOpenAI (credentials must be provided to environment):
|
||||
# client = AzureOpenAI(
|
||||
# azure_endpoint=os.getenv("AZURE_OPENAI_ENDPOINT"),
|
||||
# api_key=os.getenv("AZURE_OPENAI_KEY"),
|
||||
# api_version="2024-02-01",
|
||||
# )
|
||||
# MODEL = "sqlai" # deployment name
|
||||
|
||||
deployment_name = "sqlai"
|
||||
# Set up the OpenAI client
|
||||
MODEL = "gpt-4o"
|
||||
client = OpenAI(api_key=os.getenv("OPENAI_API_KEY"))
|
||||
|
||||
|
||||
def send_message(message: str) -> str:
|
||||
@@ -28,15 +36,60 @@ def send_message(message: str) -> str:
|
||||
"""
|
||||
system_message = """
|
||||
Du bist ein hilfsbereiter, fröhlicher Datenbankassistent.
|
||||
Du hilfst Benutzern bei der Erstellung von SQL-Abfragen für eine Datenbank eines
|
||||
großen Energieversorgungsunternehmens. Die Datenbank enthält Tabellen für Adressen,
|
||||
Zähler, Kunden und Ablesungen. Es werden Gaszähler (MeterType 'GAS') und Stromzähler
|
||||
(MeterType 'ELT')unterschieden.
|
||||
|
||||
Besonders wichtig ist, dass die Ablesungen der Werte kumulativ sind. Wenn nach dem Verbrauch
|
||||
gefragt wird, sollte der Unterschied zwischen zwei aufeinanderfolgenden Ablesungen berechnet
|
||||
werden.
|
||||
|
||||
Verwende beim Erstellen Ihrer Antworten das folgende Datenbankschema:
|
||||
|
||||
MEIN_DATENBANKSCHEMA
|
||||
CREATE TABLE Addresses (
|
||||
ID INT PRIMARY KEY IDENTITY(1,1),
|
||||
StreetName NVARCHAR(100),
|
||||
HouseNumber NVARCHAR(10),
|
||||
City NVARCHAR(50),
|
||||
PostalCode NVARCHAR(10),
|
||||
Longitude FLOAT,
|
||||
Latitude FLOAT
|
||||
);
|
||||
|
||||
CREATE TABLE Meters (
|
||||
ID INT PRIMARY KEY IDENTITY(1,1),
|
||||
Signature NVARCHAR(11),
|
||||
MeterType NVARCHAR(3),
|
||||
AddressID INT,
|
||||
FOREIGN KEY (AddressID) REFERENCES Addresses(ID)
|
||||
);
|
||||
|
||||
CREATE TABLE Customers (
|
||||
ID INT PRIMARY KEY IDENTITY(1,1),
|
||||
FirstName NVARCHAR(100),
|
||||
LastName NVARCHAR(100),
|
||||
GasMeterID INT,
|
||||
EltMeterID INT,
|
||||
FOREIGN KEY (GasMeterID) REFERENCES Meters(ID),
|
||||
FOREIGN KEY (EltMeterID) REFERENCES Meters(ID)
|
||||
);
|
||||
|
||||
CREATE TABLE Readings (
|
||||
ID INT PRIMARY KEY IDENTITY(1,1),
|
||||
CustomerID INT,
|
||||
MeterID INT,
|
||||
ReadingDate DATE,
|
||||
ReadingValue INT,
|
||||
FOREIGN KEY (CustomerID) REFERENCES Customers(ID),
|
||||
FOREIGN KEY (MeterID) REFERENCES Meters(ID)
|
||||
);
|
||||
|
||||
Füge Spaltenüberschriften in die Abfrageergebnisse ein.
|
||||
|
||||
Gib deine Antwort immer im folgenden JSON-Format an:
|
||||
|
||||
JSON FORMAT
|
||||
{ "summary": "your-summary", "query": "your-query" }
|
||||
|
||||
Gib NUR JSON aus.
|
||||
Ersetze in der vorangehenden JSON-Antwort "your-query" durch die Microsoft SQL Server Query,
|
||||
@@ -45,15 +98,20 @@ def send_message(message: str) -> str:
|
||||
Gib immer alle Spalten der Tabelle an.
|
||||
Wenn die resultierende Abfrage nicht ausführbar ist, ersetze "your-query“ durch NA, aber ersetze
|
||||
trotzdem "your-query" durch eine Zusammenfassung der Abfrage.
|
||||
Verwende KEINE MySQL-Syntax.
|
||||
Verwende KEINE MySQL-Syntax, sondern AUSSCHLIESSLICH Microsoft SQL.
|
||||
Begrenze die SQL-Abfrage immer auf 100 Zeilen.
|
||||
Formatiere den Output bestmöglich.
|
||||
"""
|
||||
system_message = "Du bist ein hilfreicher Assistent."
|
||||
|
||||
response = client.chat.completions.create(
|
||||
model=deployment_name,
|
||||
model=MODEL,
|
||||
messages=[
|
||||
{"role": "system", "content": system_message},
|
||||
{"role": "user", "content": message},
|
||||
],
|
||||
)
|
||||
return response.choices[0].message.content
|
||||
|
||||
result_str = response.choices[0].message.content.replace("```json\n", "").replace("```", "")
|
||||
if ("\n") not in result_str:
|
||||
result_str = result_str.replace("\\", "\n")
|
||||
return result_str
|
||||
|
||||
@@ -11,7 +11,7 @@ def test_db_connection() -> bool:
|
||||
This function attempts to establish a connection to an Azure SQL Database
|
||||
using the connection string stored in the environment variable
|
||||
'AZURE_SQL_CONNECTION_STRING'. It makes up to 5 attempts to connect,
|
||||
with a timeout of 240 seconds for each attempt.
|
||||
with a timeout of 480 seconds for each attempt.
|
||||
|
||||
Returns
|
||||
-------
|
||||
@@ -22,14 +22,15 @@ def test_db_connection() -> bool:
|
||||
for i in range(5):
|
||||
print(f"Trying to connect to Azure SQL Database... Attempt {i + 1}")
|
||||
try:
|
||||
pyodbc.connect(connection_string, timeout=240)
|
||||
pyodbc.connect(connection_string, timeout=480)
|
||||
print("Connected to Azure SQL Database successfully!")
|
||||
connected = True
|
||||
break
|
||||
|
||||
except pyodbc.Error as e:
|
||||
except Exception as e:
|
||||
print(f"Error connecting to Azure SQL Database: {e}")
|
||||
connected = False
|
||||
continue
|
||||
|
||||
return connected
|
||||
|
||||
|
||||
Reference in New Issue
Block a user