Database AI guide · Updated October 5, 2026
Choose a database AI tool for the work you need to verify.
Natural-language database tools can answer questions, analyze data or investigate incidents. Start with the task, the permitted access and the evidence another engineer needs.
DB-GPT: broader data analysis and applications
The official DB-GPT project describes a data assistant with database and file connections, SQL and code analysis, reusable skills and visual reports. Its scope also includes building data workflows and applications. Review its current documentation for deployment and security requirements.
OpsGraph: read-only PostgreSQL incidents
OpsGraph focuses on missing records, failed payments and stuck jobs in a private, self-hosted workspace. A model proposes SQL, while application policy checks allowed tables, read-only rules and query limits. Each investigation turn retains the query, captured rows and recorded limits.
Choose OpsGraph when you need to inspect and share PostgreSQL evidence. Choose a broader tool when your task requires other source types, charts or general code execution. OpsGraph does not integrate with DB-GPT, execute arbitrary code or automatically repair your database.
A payment incident you can reproduce
Start with OpsGraph's invented practice dataset. Ask for payment counts grouped by processor, status and error code. The expected results are 50 fastpay failures with gateway_timeout, 450 fastpay successes and 500 steadypay successes. Inspect the executed SQL before accepting the answer. These counts are a practice case, not a model accuracy benchmark.
Questions to ask before connecting a model
- Does the database login itself prevent writes?
- Which tables and records may the model receive?
- Are tenant scope, timezones, status meanings and units explicit?
- Can you inspect the SQL, captured results and query limits?
- Does the report distinguish observation from interpretation?
A citation identifies a captured result. It does not independently prove a conclusion. OpsGraph reports need human review and do not automatically redact sensitive records. The current release is designed for one trusted operator, not shared team accounts.