Follow-Up Questions Without Re-Querying

Analyse your data with your AI client · Follow-up questions on a held result

Follow-up questions without re-running the query

Your Network Agent keeps the complete result for a while. “Now break that down by region” is answered from the held result, with no second query on your database.

Personal Core and up Runs on your Network Agent 始终只读

遇到任何问题？尽管问我们。 打开支持请在“查询流”（右上角菜单 → 支持）中给我们发送消息。我们会安排专人阅读您的消息，“我不知道这个屏幕需要什么”就是一个很好的问题。

What your Network Agent does for you

Your AI client can keep asking about a result without touching your database again. Your Network Agent holds the complete result inside your network; filters, breakdowns, pages and samples are answered from it, and only each answer leaves your network. Your database ran one query.

Analysis is a conversation. Without held results, every follow-up is a new query: slower for you, more load on production, and a fresh flood of rows for your AI client. With them:

Hold: the client asks your agent to keep a result, optionally with a label. The rows it asked for still arrive, and the answer adds the held result’s id, size and expiry.
Ask again: filter it, group it, sort it or run your own SQL over it; read it page by page in the original order; or take a random sample that repeats exactly when asked again.
Totals at once: totals by group can come back in the same call that holds the result.
Combine: up to eight results held on the same agent can be joined in one follow-up.
Let go: results are released when they expire, when the client releases them, or when someone clears them in the Portal.

情况	Before	With held results
“Now by customer.”	A second query on production.	Answered from the held result.
“Show me rows 500 to 600.”	Re-run with an offset.	A page of the held result.
“Give me a random sample of the cancelled ones.”	A new query, a different random set each time.	A repeatable sample of the held result.

The plan it needs

计划	This capability	注释
免费	Not on this plan	Works as it always has, with standard results.
Personal Core, Personal Plus	包括	Part of Result analysis.
Business Core, Business Plus, Business Max	包括	Part of Result analysis.
Enterprise Core, Plus, Max	包括	Part of Result analysis.

Why it is on paid plans: your Network Agent does the heavy lifting next to your database, so the rows stay in your network and your AI client spends its attention on the thinking. That groundwork is what the paid plans add.

How long a result is held and how much each agent can hold grow with your plan; see 限制.

The Free plan works as it always has. Paid plans add the Analytic toolset. Free is a plan you keep: MCP, the REST API, federated queries and alerts, with the same results as before. Plan names and prices are on the 定价页面.

How to ask

You do not call tools yourself. Ask in plain words and your AI client picks the tool. Three asks that bring this capability in:

克劳德

ChatGPT

光标

克劳德·科德+10 more只读

Get this year’s West-region orders and keep them so we can dig in.

qs_run_query · hold

Now break that down by customer.

qs_result_query

Which results are you holding right now? Release the old one.

qs_result_windows qs_result_release

Groundwork done on your Network Agent, inside your network 已登录 MCP 活动选项卡

For people who read tool lists: hold, hold_label, group_by 和 aggregate 在 qs_run_query 和 qs_run_saved_query; qs_result_query, qs_result_page 和 qs_result_sample read a held result; qs_result_windows lists what is held and qs_result_release frees one or all. See hold on the Tools tab, 和 the follow-up query.

What comes back

克劳德

ChatGPT

光标

克劳德·科德+10 more只读

Get this year’s West-region orders and keep them handy.

qs_run_query · hold

Got them: 200,000 orders, held on your agent as w_3f9a0c12d47be801 for the next 15 minutes. Here are the first rows; ask anything about the rest.

Break the refunds down by customer.

qs_result_query

From the held result, no new query on your database:

客户	退款	已退款
C-2281	61	18,420
C-1042	44	9,760
C-0057	12	2,115

C-2281 accounts for a third of all refunds. Want its refund reasons?

Groundwork done on your Network Agent, inside your network your database ran 1 query

result_id: The held result’s id, w_ followed by 16 characters. Every follow-up names it.
expires_at: When the agent lets it go. How long results are held grows with your plan.
rows, bytes: How big the held result is.
held_refused: When the result could not be kept: the reason and, when size was the reason, how many rows would fit. The rows are still returned.

What it does not help with

A held result is a snapshot. It does not see changes made after the query ran; ask your AI client to run it again for fresh data.
Results are held on one agent for a while, then released, and an agent restart or update clears them. A newer hold can also make the agent release an older one early.
A follow-up can only read held results: one SELECT, no files, no settings, no second statement.
Holding does not combine with a profile, a quality report or a shape. Hold the rows in their own call.

限制

计划	How long a result is held	Space per agent
Personal Core, Personal Plus	15分钟	128 MB
Business Core, Business Plus, Business Max	60 minutes	512 MB
Enterprise Core, Plus, Max	up to 8 hours	4,096 MB

The space is per agent and shared by every key in your organization that holds results there.
The agent also respects the memory free on its own computer at that moment. If there is not enough, it refuses the hold and says how many rows would fit.
A follow-up stops after 30 seconds; narrow it or page the result instead.
A follow-up is shaped like any result: your token budget and cell shaping apply.

设置

Where	环境	它的作用
数据 → 数据代理, the agent’s settings	Local analysis	已启用 turns held results on or off for that agent. Budget (MB) and Hold time (minutes) default to your plan; Clear held results releases everything held there now.
MCP page, create or edit a key	Held results	Plan default, 在或离开 for the AI client using that key.
MCP page, Context tab	Held results 控制板	Lists what is held now on each agent, who holds it and how long it has left, with Release 和 Release all.

When your AI client says no

When the server refuses, your AI client receives a code and one sentence, and usually relays it. Quoted sentences are the server’s own words; the parts in italics stand for the ids a real message names.

代码	What the server says	该怎么办
`MCP_FEATURE_NOT_INCLUDED`	“Your plan does not include result analysis; profiles, data-quality reports and follow-up questions on a held result are available on Personal Core.”	Ask for rows instead, or move up to Personal Core.
`MCP_RESULT_EXPIRED`	“result w_3f9a0c12d47be801 is not held (it expired, was released, or is unknown here); call qs_result_windows for the current list, or run the query again with hold.”	Ask your AI client to run the query again. That is a new query on your database.
`MCP_RESULT_EXPIRED`	“result w_3f9a0c12d47be801 is no longer held: it was released before its time ran out, most likely to make room for a newer result on this agent; run the query again with hold (qs_result_windows lists what is still held).”	Release results you no longer need, or hold smaller ones, so the agent does not have to make room.
`MCP_WINDOW_BUDGET`	“the agent cannot keep a result of that size right now; release a held result or add a WHERE clause.” The agent’s own detail follows, with how many rows would fit.	Ask for fewer rows or columns, or ask your AI client to release a held result it no longer needs.
`MCP_AGENT_ANALYSIS_DISABLED`	“held results are switched off on agent {agent id}; a workspace admin can turn on “Local analysis” for it on the Data Agents page, or run the query without hold.”	Ask an admin to turn Local analysis back on under 数据 → 数据代理. Plain queries keep working meanwhile.
`MCP_FEATURE_NOT_INCLUDED`	“Held results are switched off for this key.”	The key’s owner edits the key on the MCP page and sets Held results 至 Plan default 或在.
`MCP_WINDOW_SQL_REJECTED`	“this statement cannot run over a held result; only one read-only SELECT over the held result’s tables is accepted (no file, catalogue or settings access, no second statement).”	Ask for one SELECT over the held result.
`MCP_WINDOW_QUERY_TIMEOUT`	“the follow-up over the held result did not finish within 30 seconds and was stopped; narrow it (a WHERE clause, fewer columns, an aggregate) or page the result.”	Ask for a narrower follow-up, or page through the result.
`MCP_AGENT_NOT_CONNECTED`	The Network Agent that owns the connector is not connected right now.	打开数据 → 数据代理 and start the agent on its computer.

示例

“Now by month, now by product, now only the cancelled ones.”

Three follow-ups, all answered from the result held after the first question. Your DBA sees one query in the logs.

“Which of this month’s customers did not order last month?”

Your AI client holds both months on the same agent and joins them in one follow-up.

还是不行？ 打开支持请在“查询流”中告诉我们您当前处于第几步以及您看到的内容。我们会帮您连接。

Facts source

Setup trouble?

Agent shows Offline, or Test connection and Browse are greyed out? Check that the Agent is online → Is the Agent online?