Data Agent and ChatBI
Many AI-related questions are not really "does the feature exist?" but "which entry should I use, which mode fits, and how should I troubleshoot?" This page focuses on those repeated boundary questions.
What is the difference between the Data Agent sidebar and ChatBI?
The key difference is not "which one is newer." It is what kind of task each one is built for.
A simple rule:
- Data Agent sidebar is better when you want to keep working inside the current page context: analysis, creation, interpretation, editing
- ChatBI is better as an independent natural-language analytics entry focused on a conversation that can keep generating charts and dashboards
If your main question is "what else can I do from this current dashboard, dataset, or page?" start with Data Agent.
If your main question is "how do I give business users a standalone NLQ entry?" start with ChatBI.
How should I choose between Agent mode, Workflow mode, and API mode?
Use this shortcut:
| Mode | Better fit |
|---|---|
| Agent mode | Open-ended questions, complex intent, the model decides the next step |
| Workflow mode | Fixed processes, predictable steps, strict execution path |
| API mode | Your own backend, multi-agent system, or workflow engine does the orchestration |
If you still have not decided between iframe, JS SDK, or API, start with Data Agent Integration Overview instead of choosing from the technical entry point alone.
If the answer quality is weak, should I tune prompts first or fix the data first?
Most of the time, fix data preparation first, then tune prompts.
A practical order:
- dataset, field, and metric naming clarity
- field/metric descriptions
- hiding fields that should not participate in Q&A
- dataset knowledge management with business terms, definitions, and synonyms
- only then add industry/private instructions, search instructions, and analysis rules
If the underlying data semantics are messy, prompts alone usually only help part of the problem.
When should I vectorize data?
When you want Data Agent to retrieve the right assets more consistently across many data assets, reduce missed recall, and improve semantic matching.
It is especially useful when:
- there are many datasets and fields
- business terminology is strong and natural language differs from field names
- long-term answer stability matters
If the current symptom is "it answers, but often chooses the wrong field or misses the business term," vectorization is often more valuable than piling on more prompt text.
When should I create analysis templates?
When users keep asking the same kind of analysis question repeatedly.
Typical examples:
- sales trend analysis
- regional performance comparison
- inventory health analysis
- weekly or monthly business review
Analysis templates are best for reusable question skeletons so users can click, tweak, and send instead of rewriting the same prompt from scratch every time.
When should I create a custom Tool?
Only when you want the Agent to know something outside the platform or perform an action.
Examples:
- query enterprise knowledge bases, FAQs, or policy documents
- query orders, inventory, ticket status, approval status
- call search services or third-party APIs
- create tickets, trigger approvals, or write back business notes
If the issue is only answer tone, welcome copy, or business phrasing, that is usually not a Tool problem. It is more likely a prompt or copy-customization problem.
What are the most important troubleshooting artifacts when Q&A fails?
Collect these first:
- the execution log from the conversation card
- the failed request response from browser
Fetch/XHR - the exported DEBUG realtime logs from system debugging
Without these three, many AI issues cannot be classified cleanly into model, retrieval, permission, data, or frontend interaction problems.
Do I need to fill the vector database address manually?
Usually no.
Complete the required services according to AI Deployment. A common mistake is not "the address is blank" but that the AI environment was never prepared correctly in the first place.
Why should I avoid switching pages or historical conversations during Agent-mode Q&A?
Because it interrupts the current Q&A flow.
If you need a stable demo or a reliable reproduction path, stay in the current page and the current conversation while testing. Otherwise user-side interruption is easily mistaken for an AI feature failure.
Further reading: