WASH_NL2SQL / app.py
hmm404's picture
Update app.py
2e19fe3 verified
raw
history blame
9.58 kB
import os
import re
import sqlite3
import warnings
import gradio as gr
import pandas as pd
from schema import schema
from langchain_nvidia_ai_endpoints import ChatNVIDIA
warnings.filterwarnings("ignore")
API_KEY = "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"
db_path = "wash_db.db"
client = ChatNVIDIA(
model="deepseek-ai/deepseek-r1",
api_key=API_KEY,
temperature=0.1,
top_p=1,
max_tokens=4096,
)
def get_table_names(schema: str):
return re.findall(r'TABLE (\w+)', schema)
def get_table_columns(schema: str, table: str):
m = re.search(rf'TABLE {table} \((.*?)\)', schema, re.DOTALL)
if m:
cols_block = m.group(1)
cols = re.findall(r'(\w+)', cols_block)
return [col for col in cols if col.lower() not in {"int", "primary", "key", "string", "bit", "real", "references"}]
return []
def agent_select_table(user_query, schema):
tables = get_table_names(schema)
# First, try longest keyword containment in table name
best = ""
best_len = 0
for table in tables:
for word in user_query.lower().split():
if word in table.lower() and len(word) > best_len:
best = table
best_len = len(word)
if best:
return best
# fallback: first table
return tables[0]
def agent_select_columns(user_query, table, schema):
columns = get_table_columns(schema, table)
selected = []
for col in columns:
if any(word in col.lower() for word in user_query.lower().split()):
selected.append(col)
return selected if selected else columns # fallback all columns
def build_sql_prompt(table, columns, schema, user_question, error_reason=None):
prompt = (
f"You are an expert SQL assistant.\n"
f"Schema: {schema}\n"
# f"Columns: {', '.join(columns)}\n"
f"User question: {user_question}\n"
"Write a valid SQLite SQL query answering the question using only the given table and columns.\n"
)
if error_reason:
prompt += f"The previous SQL query failed with the error: {error_reason}\nPlease fix and regenerate the SQL only."
return prompt
def extract_sql_query(text):
patterns = [
r"```sql\n(.*?)```",
r"```\n(.*?)```",
r"```(.*?)```",
]
for pattern in patterns:
match = re.search(pattern, text, re.DOTALL | re.IGNORECASE)
if match:
return match.group(1).strip()
# Else, look for SELECT...;
match = re.search(r"(SELECT|INSERT|UPDATE|DELETE|CREATE|DROP|ALTER).*?;", text, re.DOTALL | re.IGNORECASE)
if match:
return match.group(0).strip()
lines = text.split('\n')
sql_lines = [l for l in lines if any(k in l.upper() for k in ['SELECT', 'FROM', 'WHERE', 'INSERT', 'UPDATE', 'DELETE'])]
if sql_lines:
return ' '.join(sql_lines)
return text.strip()
def execute_sql_query(sql_query, db_path=db_path):
try:
conn = sqlite3.connect(db_path)
df = pd.read_sql_query(sql_query, conn)
conn.close()
return df, None
except Exception as e:
return None, str(e)
def summarize_with_llm(table, columns, data, user_query):
preview = data.head(5).to_markdown(index=False) if data is not None and not data.empty else "No data returned."
prompt = (
f"User query: {user_query}\n"
f"SQL result preview \n{preview}\n"
f"Summarize the result, referencing the user query and the preview.)."
)
resp = client.invoke([{"role": "user", "content": prompt}])
return getattr(resp, "content", resp) if hasattr(resp, "content") else str(resp)
# def full_pipeline(user_question):
# table = agent_select_table(user_question, schema)
# columns = agent_select_columns(user_question, table, schema)
# yield {
# table_output: gr.update(value=table),
# columns_output: gr.update(value=", ".join(columns)),
# }
# sql_prompt = build_sql_prompt(table, columns, user_question)
# sql_query, error = "", None
# # Error-handling and retry loop
# for _ in range(5):
# llm_resp = client.invoke([{"role": "user", "content": sql_prompt}])
# llm_text = getattr(llm_resp, "content", llm_resp) if hasattr(llm_resp, "content") else str(llm_resp)
# sql_query = extract_sql_query(llm_text)
# results_df, error = execute_sql_query(sql_query)
# if not error:
# break
# sql_prompt = build_sql_prompt(table, columns, user_question, error_reason=error)
# # Summarize
# summary = summarize_with_llm(table, columns, results_df, user_question)
# # Format outputs
# columns_view = ", ".join(columns)
# sql_view = f"```sql\n{sql_query}\n```"
# status_view = f"Success" if not error else f"Query error: {error}"
# out_df = results_df if results_df is not None else pd.DataFrame()
# return sql_view, status_view, summary, table, columns_view, out_df
def full_pipeline_stream(user_question):
yield "Identifying relevant table and columns...", "", "", "", "", pd.DataFrame()
table = agent_select_table(user_question, schema)
columns = agent_select_columns(user_question, table, schema)
yield f"Table '{table}' selected.", "", "", table, ", ".join(columns), pd.DataFrame()
sql_prompt = build_sql_prompt(table, columns, user_question)
sql_query, error = "", None
for _ in range(5):
yield f"Generating SQL (attempt {_+1})...", "", "", table, ", ".join(columns), pd.DataFrame()
llm_resp = client.invoke([{"role": "user", "content": sql_prompt}])
llm_text = getattr(llm_resp, "content", llm_resp) if hasattr(llm_resp, "content") else str(llm_resp)
sql_query = extract_sql_query(llm_text)
results_df, error = execute_sql_query(sql_query)
if not error:
yield f"SQL executed successfully.", f"``````", "", table, ", ".join(columns), results_df
break
sql_prompt = build_sql_prompt(table, columns, user_question, error_reason=error)
yield f"Retrying due to error: {error}", f"``````", "", table, ", ".join(columns), pd.DataFrame()
if not error:
summary = summarize_with_llm(table, columns, results_df, user_question)
yield "Summarization complete.", f"``````", summary, table, ", ".join(columns), results_df
else:
yield f"Final error: {error}", f"``````", "No summary due to error.", table, ", ".join(columns), pd.DataFrame()
def full_pipeline(user_question):
# Step 1: Identify table and columns first
# yield "", "", "", "", "", pd.DataFrame()
table = agent_select_table(user_question, schema)
columns = agent_select_columns(user_question, table, schema)
# Immediately return only these two visible outputs
yield {
table_output: gr.update(value=table),
columns_output: gr.update(value=", ".join(columns)),
}
# Step 2: Continue with downstream pipeline
sql_prompt = build_sql_prompt(table, columns, schema, user_question)
sql_query, error = "", None
for _ in range(5):
llm_resp = client.invoke([{"role": "user", "content": sql_prompt}])
llm_text = getattr(llm_resp, "content", llm_resp) if hasattr(llm_resp, "content") else str(llm_resp)
sql_query = extract_sql_query(llm_text)
results_df, error = execute_sql_query(sql_query)
if not error:
break
sql_prompt = build_sql_prompt(table, columns, schema, user_question, error_reason=error)
sql_view = f"\n{sql_query.strip()}\n"
status_view = "Success" if not error else f"Query error: {error}"
out_df = results_df if results_df is not None else pd.DataFrame()
yield {
sql_output: gr.update(value=sql_view),
status_output: gr.update(value=status_view),
results_output: gr.update(value=out_df)
}
summary = summarize_with_llm(table, columns, results_df, user_question).strip()
yield {
# sql_output: gr.update(value=sql_view),
summary_output: gr.update(value=summary),
}
with gr.Blocks(title="NL2SQL Pipeline)") as gradio_interface:
gr.Markdown("## NL2SQL Pipeline ")
gr.Markdown("Enter a question about the water supply database. The agent will select relevant table/columns, generate and retry SQL on error, show results and a grounded summary.")
with gr.Row():
input_text = gr.Textbox(label="Enter your natural language question", lines=3)
with gr.Row():
submit_btn = gr.Button("Generate, Execute & Summarize", variant="primary")
with gr.Row():
table_output = gr.Textbox(label="Table Used", lines=1)
columns_output = gr.Textbox(label="Columns Used", lines=2)
with gr.Row():
sql_output = gr.Textbox(label="Generated SQL Query", lines=5)
with gr.Row():
status_output = gr.Textbox(label="Execution Status", lines=2)
with gr.Row():
results_output = gr.Dataframe(label="Query Results", interactive=False)
with gr.Row():
summary_output = gr.Textbox(label="LLM-Grounded Summary", lines=5)
with gr.Row():
abort_btn = gr.Button("Abort / Stop Task")
running_event=submit_btn.click(
fn=full_pipeline,
inputs=input_text,
outputs=[sql_output, status_output, summary_output, table_output, columns_output, results_output]
)
abort_btn.click(
None,
inputs=None,
outputs=None,
cancels=[running_event],
queue=False
)
if __name__ == "__main__":
gradio_interface.launch()