Skip to content

Part 8: NL2SQL Agent with the IBM CE Accelerator

Outcome: An editable draft agent that answers natural-language questions against the workshop PostgreSQL database.

This is a mandatory standalone use case. It does not rename or modify the customer-support system from Parts 2–7.

What we will implement

You will discover the workshop PostgreSQL schema, generate an editable natural- language-to-SQL draft agent, validate and deploy it, then test questions against the discovered tables and metadata.

The accelerator performs database discovery and initial agent generation in its own UI. Bob resumes as the development interface after the generated agent is available in the participant project or Orchestrate environment.

PostgreSQL → Data discovery → NL2SQL configuration → Orchestrate draft agent

1. Open the accelerator

Open the verified IBM CE NL2SQL Accelerator.

Use only the shared database connection supplied by the instructor. Do not enter personal or production database credentials during the workshop.

2. Run data discovery

Under Database Connection, keep the pre-filled workshop PostgreSQL values. Start discovery and wait for it to complete.

Review the discovered tables, relationships, and descriptions. Correct unclear table or column descriptions before generating the agent; this metadata affects SQL quality.

3. Connect watsonx Orchestrate

Under WXO Settings, enter:

  • the workshop watsonx Orchestrate instance URL
  • your IBM Cloud API key
  • environment draft

Treat the API key like a password. Enter it only in the accelerator's WXO settings form.

Select Generate Agent.

NL2SQL WXO settings

Warning

Do not continue while WXO Settings Incomplete or another configuration warning remains. Save or re-enter the settings and regenerate the summary.

4. Validate and deploy the draft

Review the generated agent name, tools, and warnings.

Generated NL2SQL agent summary

Select Deploy, then:

  1. Confirm the instance URL.
  2. Select Validate.
  3. Continue only after validation succeeds.
  4. Select Deploy Agent.

NL2SQL deployment dialog

5. Test the agent

Open the generated agent in watsonx Orchestrate and try:

  1. What were total units sold and net revenue on January 1, 2024?
  2. Which month in 2024 had the highest net revenue?
  3. How many active stores are open on Shabbat?

Check that:

  • the generated SQL uses the intended tables and filters
  • totals and dates are plausible
  • the agent explains when a question cannot be answered from the discovered schema
  • no unrestricted or sensitive table is exposed

The result may include a chart:

6. Continue with Bob

The generated agent is editable. In Bob, ask:

Inspect the generated NL2SQL agent without changing it. Summarize its tools,
data-access boundaries, and current safeguards. Recommend one small improvement
and cite the ADK documentation that supports it.

Choose one recommendation, then ask:

Implement the approved improvement in the generated NL2SQL project. Preserve
its existing database tools and data-access boundaries. Validate every changed
artifact, import or re-import it into draft using the existing .venv, and
suggest test prompts for a valid query, an unanswerable question, restricted
data, and ambiguous business language. Show me the changes and import results.

Possible improvements include guidelines for allowed query types, clearer refusals for restricted data, or a collaborator that explains results using approved business documentation.

Checkpoint

  • Discovery completed before agent generation
  • Metadata was reviewed
  • WXO validation succeeded
  • The agent was created in draft
  • Three sample questions were checked against the database schema
  • Data-access limitations are understood
  • Bob reviewed and tested one approved improvement

You have completed the mandatory workshop path.