Back office ops · Production

Woowa Brothers BADA team builds QueryAnswerBird LLM-based AI data analyst with Data Discovery feature

The problem

Employees at Woowa Brothers with varying data literacy spent significant time understanding complex queries, and non-data roles such as product managers lacked effective ways to explore internal data beyond text-to-SQL query generation.

First attempt

The initial text-to-SQL feature had gaps in table selection accuracy and business logic application; users frequently skipped the user guide and asked vague questions, receiving unsatisfactory answers; and LLM-generated table metadata suffered from hallucination errors.

Workflow diagram · grounded in source
1
User submits question
trigger
“A Single-Turn conversation consists of a single question and response.”
2
Router Supervisor classifies intent
ai_action
“we implemented a Router Supervisor chain that accurately understands user intent and delivers appropriate answers”
3
Question quality scored and gated
validation
“Questions that meet a certain score threshold and pass the classification model proceed to the next stage—information acquisition. Questions that do not meet these criteria are automatically prompted with a message to ask more specific q…”
4
Vector store retrieves catalog and log data
ai_action
“This feature utilizes a vector store to manage data from the company's Data Discovery Platform, including the Data Catalog, which contains detailed information on all internal tables and columns and the Log Checker which manages user beh…”
5
LLM interprets data and generates response
ai_action
“This prompt utilizes the Plan and Solve Prompting method, developed to overcome the limitations of the existing chain-of-thought (CoT) methodology, to guide the interpretation of queries and tables.”
6
Standardized response delivered to user
output
“provides responses to users in a standardized output format”
Reported outcome

QAB's Data Discovery feature improved user satisfaction and reliability across data and non-data roles, with the team expecting QAB to significantly boost internal productivity.

Reported metrics
User satisfaction and reliabilityimproved user satisfaction and reliability
Internal productivitysignificantly boosting internal productivity
Off-topic question rate (diagnostic finding)more than 10%
Reported stack
SQLGlotLangGraphSlack
Source
https://tech.deliveryhero.com/blog/introducing-the-ai-data-analyst-queryanswerbird-part-2-data-discovery/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

QAB's Data Discovery feature improved user satisfaction and reliability across data and non-data roles, with the team expecting QAB to significantly boost internal productivity.

What tools did this team use?

SQLGlot, LangGraph, Slack.

What results were reported?

User satisfaction and reliability: improved user satisfaction and reliability; Internal productivity: significantly boosting internal productivity; Off-topic question rate (diagnostic finding): more than 10% (source-reported, not independently verified).

What failed first in this deployment?

The initial text-to-SQL feature had gaps in table selection accuracy and business logic application; users frequently skipped the user guide and asked vague questions, receiving unsatisfactory answers; and LLM-generat…

How is this back office ops AI workflow structured?

User submits question → Router Supervisor classifies intent → Question quality scored and gated → Vector store retrieves catalog and log data → LLM interprets data and generates response → Standardized response delivered to user.

Related back office ops cases