This guide assumes you have connected the Lovrabet runtime MCP. Ask questions the way shown below to query data, run statistics, execute existing SQL / business capabilities, and perform controlled writes.

If you haven't configured it yet, complete <cite doc-id="J3pxwfjAmiZfzhkaFHbc5rFSnUd" file-type="wiki" title="接入 Lovrabet 运行态 MCP" type="doc"></cite> first.

Select the right app first

With a multi-app key, confirm the current app before querying data:

列出我能访问的应用，并告诉我当前会话选中的是哪一个。

切换到「<应用名称>」，之后都在这个应用里操作。

Find datasets and fields

When you're not sure of the dataset name:

帮我推荐和「订单」相关的数据集，并说明为什么推荐。

Check the fields before querying:

查看「<数据集名称>」的字段、类型和可用操作。

Query and aggregate business data

Filtered query

在当前应用里查询客户资料：城市等于北京，只返回公司名称和联系人，按创建时间倒序，先看前 20 条。

Fetch a single record

根据主键 <记录ID> 取出这条客户记录的完整信息。

Aggregation

按城市统计客户数量，并按数量从高到低排序。

A query is working when you see structured results — lists, fields, row counts — rather than a vague summary.

Run existing SQL and business capabilities

See what SQL is available

列出当前应用里已保存的自定义 SQL，并简要说明各自用途。

Review details, then execute

先查看 SQL「<SQL名称或编码>」的内容和参数，确认无误后再执行，参数为：<参数说明>。

Run an existing Backend Function

先查看 BFF「<函数名>」的说明，再带参数执行：<参数说明>。

Preview before writing

For create and update operations, ask for a preview first and confirm the fields before anything is actually written:

先预览创建一条客户记录，不要真正写入。字段：公司名称=示例公司，城市=上海。

确认刚才的预览没问题，请真正创建这条客户记录。

Deletion is a high-risk operation and requires explicit confirmation. You can say:

删除客户记录 <记录ID>。我已核对过目标记录，请确认后执行删除。

If the assistant asks you to confirm again, verify the app, dataset, and primary key before replying.

File uploads and OCR

The runtime MCP does not read local file paths. To upload, hand the file content to the assistant (for example, as a chat attachment or content you provide), and the MCP processes it as uploaded content.

把我刚提供的发票图片做 OCR 识别，场景用发票，识别后把关键字段整理成表格。

上传我提供的附件到当前应用，并返回可用于后续引用的文件标识。

Knowledge base and personal draft capabilities

在当前应用知识库里搜索「退货规则」，给出最相关的几段原文摘要。

把下面这段说明保存为我的个人知识条目，标题为「<标题>」：
<正文>

Running personal scripts is also a high-risk operation: review the content and give explicit confirmation before executing.

App configuration keys (no plaintext)

You can check whether a configuration key exists and whether it has a value, but the MCP never returns configuration plaintext:

检查当前应用是否配置了键「<配置键名>」，只告诉我是否存在、是否有值，不要输出密钥内容。

Completion checklist

• [ ] The current session is operating on the correct app

• [ ] Queries returned real business data or a definitive empty result

• [ ] Writes were previewed or field-checked beforehand

• [ ] High-risk operations (deletes, personal scripts) were double-confirmed

Troubleshooting

Field names in the results are hard to read

Have the assistant read the dataset field descriptions first, then ask in business terms. If a field's display name itself is miscalibrated, correct it in the dataset on the Lovrabet workspace before querying.

Rate-limit prompts or "try again later" messages

Too many requests in a short window trigger rate limiting. Wait as the response instructs, then retry only the failed step — don't resubmit the entire flow in one go.

The assistant only offers advice without calling the MCP

Explicitly ask it to "use the Lovrabet runtime MCP to query real data". If that still doesn't work, go back to the connection guide and check whether the MCP is enabled and authenticated.

What not to expect from the runtime MCP

Project initialization, database analysis, standard page generation, menu publishing, and local config file maintenance are not part of the runtime MCP. Do these in natural language through the Rabetbase development suite.