Ask ConnectGain

Overview

Ask ConnectGain (the button in the dashboard top bar) answers three different kinds of question:

1. Product questions — how a feature works, how to set something up, what a setting does, why the product behaves a certain way. Answers come from the official documentation at docs.connectgain.cloud, and the assistant links the page it used so you can read further.
2. Technical support questions — something is not working in your workspace right now. The assistant checks your live configuration and data rather than the documentation, and tells you what it found. See Troubleshooting with the assistant.
3. Workspace data questions — figures from your own records, such as deal counts for a period or which team members have been most active.

Who can use it

Ask ConnectGain is part of the paid plans. On the free plan the button is still there and still opens — it shows what the assistant does and a link to the plans — but it will not answer questions. It is deliberately visible rather than hidden: a feature nobody can see is a feature nobody knows they are missing.

The public demo workspace keeps full access, so you can try it before buying.

Within a paid workspace, access depends on the role:

Capability	Owner	Admin	Agent / other roles
Product questions from the documentation	Yes	Yes	Yes
Troubleshooting your workspace	Yes	Yes	Yes, narrowed to their own view
Workspace data (deals, tasks, contacts, inbox, analytics)	Yes	Yes	No

Everyone can troubleshoot. The person who notices that chats are missing is almost always the agent looking at the inbox, and the most common true answer — "that conversation is assigned to a colleague, and a setting on your account hides it from you" — is one they can be told immediately instead of an hour later, second-hand, by someone who was not there.

What "narrowed" means

For an agent, a manager or a developer, a diagnostic answers from their own seat. One rule decides where the line falls: it never shows them anything they could not already find by clicking around the app.

So colleagues' names stay — everyone already sees the full team in the assignee picker, and "assigned to a colleague" without saying which one turns a solvable problem into a scavenger hunt. What is left out is anything behind the Team settings screen they cannot open: colleagues' permissions, their channel access, who is online and how much each is carrying. Those come back as counts instead — "4 of 7 people can currently receive conversations" — which is enough to understand the situation without reading anyone's settings.

Conversations outside their reach are reported as existing, on which channel, and why they cannot see them — but not their contents. That is deliberately the shape of the answer they need: the chats are not missing, they are over there, and this is the setting in the way.

A narrowed answer always says it is narrowed, and when the cause genuinely sits somewhere only an owner or admin can look, the assistant says exactly what to ask them to check — one sentence that can be forwarded as it is, rather than "ask your admin".

Owners and admins see all of it, including any specific person's view: ask about a colleague by name and the assistant will check what that person can see.

The assistant can be hidden for the whole organization from Settings → Organization → Hide AI assistant.

How answers are grounded

The assistant does not answer product questions from memory. It searches the documentation first and answers only from the passages it finds. If nothing relevant comes back, it says so and suggests contacting support rather than guessing.

Retrieval is hybrid — keyword search plus semantic (vector) search:

• Keyword search catches exact terms: a setting name, an error code, an API key prefix.
• Semantic search catches rephrasings, and lets a question asked in Arabic (or any of the other supported UI languages) match the English documentation.

You can ask in any language the product supports. The assistant searches the English documentation and replies in the language you wrote in.

Troubleshooting with the assistant

A product question asks how something works. A troubleshooting question reports that something is wrong right now — chats that are not there, messages that stopped, conversations that are not reaching anyone. The documentation cannot answer the second kind, because the answer is in your own workspace. For these the assistant reads your live configuration and data instead, and tells you what it found rather than what the product does in general.

What it can check

One customer's chat history. Give it a phone number, an email address, a contact name or a platform thread id and it reports every contact card that matches, every conversation on every channel, how many messages each holds, and when each last received and last sent something.

It also normalizes the number the way the channels do, so a number saved as +20… on one card and 0… on another is recognized as the same person. That matters, because a second contact card holding half the history is the most common reason a chat looks missing — nothing was lost, the history is split across two cards and the agent opened one of them. Merging the contacts puts it back together.

Whether a specific person can see a conversation. This is checked automatically for whoever is asking. As an owner or admin you can also name a colleague, and the assistant checks each conversation against what that person is allowed to see: whether it is assigned to a colleague while their "see all conversations" permission is off, whether they were granted the channel it is on, and whether it is closed while they are looking at the open queue.

This is the second most common answer, and it is invisible from the reporter's own screen: from where they sit, a conversation they are not allowed to see and a conversation that does not exist look exactly the same.

Whether messages are still arriving. Per channel, it compares how much arrived recently against the same length of time immediately before, so a channel that has stopped is distinguishable from one that was always quiet. It reports expired connections with the date they expired — which is the date messages stopped — messages that failed to send, and incoming messages that reached ConnectGain and then failed to be saved. Those last ones are the only case where messages really are missing from the inbox despite being sent and delivered; they retry automatically.

The Inbox Health page in the app (the health icon at the top of the Inbox) shows the same ground live if you would rather watch it than ask.

Why conversations are not reaching agents. It reports how distribution is configured, how many conversations are sitting unassigned right now and for how long, and the specific reason each team member is or is not eligible to receive one — a permission that is off, a channel they were never granted, being outside their shift, or holding a manager role.

Inbox counts for a period you name. Owners and admins can ask for the conversation numbers over a window — today, yesterday, this week, last week, this month, last month, the last 7 days or the last 30. Each is a whole calendar window with a start and an end, worked out in the workspace's own timezone, so "yesterday" asked at 1am in Cairo means yesterday in Cairo rather than the day before that.

The breakdown names every status — open, closed and pending — alongside the unassigned count and your own. The figures are exact counts rather than a sample, so they stay right in workspaces with more than a thousand conversations.

What it will not do

• It never reads message content. Every check above works from counts, timestamps and settings. The text of your customers' conversations is not needed to answer "did it arrive and can you see it", so it is never read and never sent to the AI provider.
• It will not tell you data was deleted. Nothing in ConnectGain removes a conversation on its own. If a trace comes back empty the message either never arrived or is stored under a different identifier, and the assistant says which rather than guessing.
• It will not tell you the product is at fault. See below — this one is a firm rule, not a tendency.

When it cannot explain what you are seeing

The assistant will never tell you that ConnectGain has a bug, and it will never hedge its way there either — no "this looks like a defect", no "that shouldn't happen".

That is deliberate, and it is not about appearances. The assistant reasons from a handful of counts, timestamps and settings, which is nowhere near enough to attribute a fault to the product. Guessing wrong in that direction is expensive in both directions: it sends engineers hunting something that is not there, and it tells you to stop looking and wait for a fix nobody was asked to make.

So when its checks come back clean and what you are describing is still real, it routes the question to a person instead. It tells you it is doing so, and sends the support team your workspace name, your name, everything you asked in that chat, and what it already checked — so whoever picks it up starts with the full picture instead of asking you to explain it again.

It also opens a ticket for the issue at the same time, so it is tracked rather than scrolling away in a chat window, and the alert the team sees carries the ticket number. Nobody has to remember to raise one.

You do not need to do anything to trigger this. It happens on its own whenever the assistant reaches the end of what it can verify. You can also just ask for a human and it will do the same — every starter prompt list ends with a Talk to a person option for exactly that.

This is separate from the Chat with our support team button under the chat, which opens WhatsApp and puts you in touch directly. Use that when you want to talk to someone now; the automatic routing above is for when the assistant has already looked and come up short.

If the assistant itself fails

Software falls over. When Ask ConnectGain does, it does not tell you something went wrong and ask you to try again — that would contradict everything else it says, and it would send you round a loop that fails the same way while nobody is told it failed at all.

Instead you get the same reply as any other escalation: the issue is going to a person. Behind it, the failure is reported to the engineering team with your workspace, your name, the question you had just asked and a link that reopens it, so it can be reproduced without asking you to describe it.

Repeated failures of the same kind are grouped together, so a problem that hits several people in an hour arrives as one thing to fix rather than a hundred copies of it.

If the answer is simply wrong

Every answer has a thumbs-down under it. Pressing it files a bug against the product with the question you asked, the answer you were given and which workspace saw it — which is everything needed to reproduce it. You do not have to explain what was wrong, though there is room to say so if you want to.

It is a bug rather than a support ticket on purpose: a wrong answer is a defect in this product, not a question for the support desk, and it belongs on the same board as the crashes.

Repeat presses on the same question are grouped into one bug, so a widely-seen bad answer arrives as one thing to fix rather than one per person who hit it.

What it will still tell you plainly

Not blaming the product is not the same as being vague. Facts from your own workspace are still reported directly — that a connection expired on a particular date and nothing has arrived since, that incoming messages failed to save and are retrying, that conversations are assigned to a colleague and a permission is switched off. The line is between telling you what your workspace shows, which it does, and telling you why in a way that blames the product, which is support's call.

Getting a useful answer faster

Have the number or email of the customer whose chats are involved — that one detail changes the answer more than anything else.

If you are an owner or admin asking on someone else's behalf, name them too. Without a name the assistant checks your own view, which will not show you why they cannot see something.

Opening it on a specific question

Ask ConnectGain can be opened with a question already typed in, by adding ?ask= and the question to your dashboard address:

https://dashboard.connectgain.cloud/?ask=Why%20did%20WhatsApp%20stop%20receiving

The assistant opens with that question waiting in the box. It does not send it — you press send, so a link someone shares with you can never ask something on your behalf or spend your workspace's AI credit without you.

This is useful in a few places:

• Support sending you back to it. When the team resolves something, they can send a link that reopens the exact question so you can see the answer in context rather than rereading a chat thread.
• Your own team. Put one in an internal runbook or a pinned message — "if conversations stop arriving, start here" — so people land on the right question instead of describing the problem from scratch.
• Reproducing what someone saw. A colleague can send you the link rather than a screenshot of the reply.

The parameter disappears from the address bar as soon as the assistant opens, so refreshing the page will not reopen it over whatever you have since typed.

What it knows

The knowledge base covers the public documentation: getting started, the user guide, the admin guide, use cases, integrations, the REST API, webhooks, and the reference section (FAQ, glossary, troubleshooting, best practices).

It deliberately excludes internal engineering material — developer docs, architecture notes, runbooks, audits and archived pages — so it will not surface implementation detail or stale superseded guides.

Keeping it current

The knowledge base rebuilds automatically whenever documentation changes land on the main branch. A GitHub Actions workflow re-chunks the docs and re-embeds only the passages that actually changed, so routine edits are cheap and the assistant is never more than one deploy behind the published docs.

Adding a page to exclude_docs in docs-site/mkdocs.yml removes it from the public site and from the assistant's knowledge base — one list governs both.

Requirements

The assistant needs your organization's Google Gemini API key, configured in Settings → AI Provider, to write its replies. Without a key you will see a prompt to add one. See BYOK Gemini setup.

Searching the documentation itself does not use your key — that runs on ConnectGain's own key, so documentation search behaves identically for every workspace.

When the assistant itself misbehaves

"I couldn't find that in the documentation." The topic may genuinely not be documented, or may be phrased very differently. Try the product's own vocabulary — "sequence enrollment" rather than "drip list", "bot flow node" rather than "chatbot step".

Answers cite a page that looks out of date. The assistant mirrors the published documentation. If the page itself is wrong, correcting the page corrects the assistant on the next deploy.

The assistant only shows an upgrade panel. The workspace is on the free plan. Ask ConnectGain is included from the first paid tier upwards; See plans in that panel opens the pricing page.

An agent says the assistant refuses their question. Check whether they asked for workspace data — deals, tasks, contacts or analytics figures — which is restricted to owners and admins by design. Troubleshooting is not: everyone can run it, narrowed to their own view.

A trace reports nothing for a number you are certain messaged you. Check the number itself first: one saved without its country code, or a customer writing from a second number, produces exactly this. If the number is right, ask about message delivery for that channel instead — the answer is then about the channel, not the customer.

The assistant says a conversation exists but the agent still cannot find it. Ask again naming that agent. Without a name it can only report that the conversation is there; with one it checks that person's permissions and channel access and tells you which filter is hiding it.

It said it was routing the issue to support — what happens next? The support team receives it with your workspace, your name, the whole chat and what the assistant already checked. Nobody needs to re-explain it. If you want to talk to someone straight away rather than wait, use the Chat with our support team button under the chat as well.

An agent says the assistant told them to ask an admin. It should not, before it has looked. Everyone can run the diagnostics, and the commonest cause of missing chats is a setting on the asker's own account — which a narrowed check shows them directly. An admin is only needed when the cause genuinely sits somewhere they cannot see, and then the assistant should say precisely what to ask for rather than sending them away empty-handed.