Challenge 7: Conversational analytics for insights
Previous Challenge Next Challenge
| Target Persona: Data Analyst / Product Manager | Estimated Duration: 30 minutes |
Introduction
Disneyland park managers need to query this complex multi-silo dataset (reviews, wait times, graph movements, classifications) without writing SQL. In this challenge, you will build a data agent using BigQuery’s Conversational Analytics. You will leverage the context layer defined in the Knowledge Catalog.
Description
Task 7.1: Initialize the Conversational Analytics Agent
- In BigQuery Studio, navigate to the Agents tab.
- Create a new agent named
disney_park_analystand connect it to the table under disney dataset. You can also put the previously created BQ graph as a knowledge source. (You can either choose tables or a Graph, not both)
Task 7.2: Use the Knowledge Catalog
To prevent the agent from hallucinating, you can leverage the previously configured Knowledge Catalog.
- Metadata Descriptions: Choose sources with curated metadata.
- Synonyms & Vocabulary: Make sure business terms are imported.
Task 7.3: Define Golden Queries
Train the agent’s SQL generation engine by providing Golden Queries—pre-approved, highly accurate SQL templates that the model can reference.
Provide golden queries for:
- Joining the attractions table with the wait-time forecasts.
- Querying the graph routing table.
Task 7.4: Execute Multi-Silo Prompts
Once configured, test the agent in the chat interface. Ask complex, cross-dataset questions like:
- « Which attractions have the highest negative sentiment today, and what is the most common path visitors take after leaving them? »
Success Criteria
To validate this challenge, you must demonstrate the following:
- Show the Conversational Analytics agent
disney_park_analystconfigured in the BigQuery Console. - List the synonyms and Golden Queries you defined in the agent’s configuration.
- Show a screenshot or proof of the chat interface successfully answering the complex multi-silo prompt without any SQL syntax errors.