Knowledge Hub
An AI chat widget that answers visitors' questions strictly from the knowledge you give it. No made-up answers - it only speaks from the documents in your Knowledge Base.
Overview
Upload your documents to a Knowledge Base, and the chatbot answers questions from them with citations, scoped to your workspace only:
- Add documents, spreadsheets, presentations, email, images, audio, or video to a Knowledge Base.
- The chatbot answers using that content and shows where each answer came from.
- Answers are limited to your workspace - your data is never mixed with anyone else's.
Build your Knowledge Base
Upload your files, then group them into collections (for example "Support docs" or "Product FAQ"). Collections let you decide exactly which knowledge a given chat widget can draw from.
What you can upload
Text is extracted from all of these, so the chatbot can quote them in answers:
- Documents - PDF, Word (.docx and the older .doc), OpenDocument (.odt), RTF, EPUB.
- Spreadsheets - Excel (.xlsx and the older .xls), OpenDocument (.ods), CSV, TSV, dBase.
- Presentations - PowerPoint (.pptx and the older .ppt), OpenDocument (.odp).
- Email - .eml, Outlook .msg, and .mbox archives.
- Text and code - .txt, Markdown, HTML, XML, JSON, YAML, TOML, SQL, subtitles (.srt, .vtt), calendar (.ics), contacts (.vcf), and common source files.
- Images - PNG, JPEG, WebP, TIFF, BMP, GIF. Text is read out of the picture using AI OCR.
- Audio - MP3, WAV, M4A, AAC, FLAC, OGG. Speech is transcribed using AI.
- Video - MP4, MOV, WebM, AVI, MKV, WMV, and more. The spoken audio is transcribed; text shown on screen is not read, and a video with no sound produces nothing.
AI OCR and AI transcription can each be switched off in Knowledge Hub settings. When they are off, images, audio, and video still upload but are left as Stored.
When you create a knowledge base you choose:
- Who can see it - only you, your whole team, or specific roles and members.
- Embedding model - keep the workspace default or pick a different one for this knowledge base.
- Vector database - keep the workspace default or, on an entitled plan, store this knowledge base in its own Qdrant index.
Each knowledge base keeps its own separate index, so different knowledge bases can use different embedding models and vector databases without affecting one another. You can change any of these later from the knowledge base page.
Add the chat widget
- Go to Knowledge Hub → Developer and open Embed the widget.
- Under Widget design, pick at least one collection under Knowledge base, then fill in Name and Greeting. You can also add a bubble tagline and up to six Quick questions, and set the colours. The preview beside the form shows the result.
- Under Public widget keys, type your website address as the Allowed origin (for example
https://www.example.com) and click Generate key. Copy the key from the Save this key now window. - Paste this into your site, just before
</body>.data-mode="floating"puts the chat bubble in the bottom-right corner:
<script src="https://app.softsolz.uk/softsolz.js"
data-service-id="smart-chat"
data-public-key="pk_live_your_key"
data-mode="floating"></script>
The widget answers only from the collections chosen for that public key. To change them, or the name, greeting and colours later, use Edit widget on the key.
Read what visitors asked
Every conversation the widget has on your website is saved. Open Website Chat under Knowledge Hub to read them.
The list shows what each visitor asked first, which site they were on, how many messages were exchanged and whether you have read it. Open one to see the whole exchange, question and answer.
Two things worth knowing:
- A visitor who comes back later carries on the same conversation rather than starting a new one, so you see the whole story in one place.
- Close a conversation once you have dealt with it. Closed ones drop out of the default view but are never deleted unless you delete them.
This is also the honest answer to "what do people actually ask us?" - it is usually the fastest way to find the gaps in your Knowledge Base.
Conversations are kept for 180 days and then removed automatically. Reading them needs the Read conversations from the website chat widget permission, and deleting one needs Delete a website chat conversation.
AI Assistant
With Knowledge Hub installed, the AI Assistant sits in the sidebar under Dashboard. It answers questions about your workspace, and it can also create and edit records for you. You can give it a name of your own.
Let it make changes
Ask for a change in plain words, for example "Create a form called Contact with name, email and message fields". The Assistant makes it straight away and you stay in the chat. Its reply shows a card for what it made:
- Open takes you to the record, with a link back to your conversation.
- Undo reverses the change. It works for 24 hours, only for the person who asked, and stops if someone has edited the record since, so their work is never lost.
- The card shows the record's status and what changed, and offers next steps such as Publish.
Follow up with "make email required" and a second card shows the edit. An answer that made changes cannot be regenerated, so the same change is never made twice.
What it can change: records in the services you have installed, for example forms and their fields, blog and social post drafts, contacts, tasks, booking links, campaigns and knowledge bases. See Build forms by asking for a worked example.
- It works from the Quick and All data scopes, or with a service selected. In the Knowledge scope it only answers from your documents.
- It never publishes, sends, pays, refunds or deletes on its own. Those become a request for a person to approve.
- A change to something already live, such as a published form, also waits for approval.
- It can only do what your role allows, and it asks before creating more than three new things from one message.
Review rules
Choose which AI changes a person must approve first. Open AI Review → Settings and turn Require approval on or off for each action. Changing these needs the AI Review manage permission.
- Publishing, sending, paying and deleting always need approval. With Require approval off you can approve them from the card in the chat; with it on they wait in AI Review → Pending.
- Creating and editing drafts happen straight away with Undo while Require approval is off. Turn it on and those changes wait in AI Review too.
Every approval and rejection is kept in AI Review → History, and every change the AI makes, including an undo, is recorded in your workspace audit log.
AI Agents
Agents are reusable personas (for example an "AR Assistant") that you can chat with inside the dashboard. Each one carries its own instructions and knowledge so your team can pick the right helper for the task at hand.
Activity & Personalization
Open Knowledge Hub → Activity to see your own usage as a contribution heatmap - active days, total chats, and your current streak - so you can tell at a glance how much you lean on the assistant.
The Personalization tab lets the assistant learn how you like to work and tailor its answers over time. It is a single switch:
- On - the assistant quietly learns your preferences (tone, format, the topics you care about) from your chats and applies them to future answers.
- Off - it stops learning and pauses everything it has stored. Nothing is used until you turn it back on.
You can also tell the assistant to "remember" something directly in chat, or use the Remember this about me action on any answer. Everything the assistant keeps is listed on this tab, so you can review, edit, or delete any item - or clear it all. It is your personal memory: only you see it, and it is included in your data export and removed if you erase your account.
Improve your answers
Once documents are in, nine settings change how well the Hub finds and answers from them - teaching it your company shorthand, matching how each knowledge base is split up and searched to what is in it, and stopping out-of-date documents being quoted. Every one is off until you turn it on.
Knowledge Hub settings - all nine, step by step
Test it step by step
Follow these steps once from start to finish to prove Knowledge Hub works for your workspace. You need nothing outside SoftSolz. Reading a document and answering questions use AI credits from your plan, so keep the test small.
- Get access. You need the Super Admin or Admin role, or a custom role that includes the Knowledge Hub permissions (Members & access → Roles).
- Pick where to test. To try things without touching live data, open the workspace menu and choose Switch to Sandbox mode. The sandbox has its own knowledge bases, keys and conversations.
- Install Knowledge Hub. Open Marketplace, find Knowledge Hub and click Install. Knowledge Hub and the AI Assistant now appear in the sidebar.
- Create a knowledge base. Go to Knowledge Hub → Ingestion, click New knowledge base, name it "Test FAQ", choose Whole team and click Create.
- Add one document. On the same page click Add text, pick "Test FAQ", give it the title "Refund policy" and the text "Customers can return any item within 30 days of delivery for a full refund." Click Save. After a few moments the document shows as Ready under Knowledge Hub → Library.
- Ask the Assistant. Open AI Assistant, switch to the Knowledge scope and ask "How long do customers have to return something?". The answer says 30 days and shows "Refund policy" as its source.
- Set up the widget. Go to Knowledge Hub → Developer and open Embed the widget. Under Knowledge base pick "Test FAQ", and fill in Name (for example "Help") and Greeting. Enter
https://developer.softsolz.ukas the Allowed origin and click Generate key. Copy the key from the Save this key now window. - Try the chat in the Widget Tester. Open developer.softsolz.uk/widget-tester and choose Knowledge Hub. In the Embed snippet box, replace
pk_live_REPLACE_WITH_YOUR_KEYwith your key and click Load widget. A chat bubble appears at the bottom right. Open it and ask "Can I get a refund?": the reply comes from your Refund policy. - Read the conversation. Back in SoftSolz, open Knowledge Hub → Website Chat. Your test conversation is listed with the question you asked first. Open it to see both sides.
- Put it on your own site. On the same Developer page, generate a second key with your own website address as the allowed origin (
wwwand nowwwcount as different addresses). Paste the snippet from Add the chat widget into your site, publish the page and ask one question. - Tidy up. Delete the test document or knowledge base, close the test conversation, and revoke any key you no longer need.
For developers
In the sandbox, open Knowledge Hub → Developer → Use the API, click New API key and tick View Knowledge Base. Sandbox keys start with sk_test_. Then list your knowledge bases from your server:
curl https://app.softsolz.uk/api/v1/services/smart-chat/kb/collections \
-H "Authorization: Bearer sk_test_your_key"
A 200 reply lists each knowledge base the key can see, with how many documents it holds. A plain API key only sees knowledge bases shared with the whole team. Sending a chat message through the API uses AI credits, the same as asking in the dashboard. See the API Reference and the Knowledge Hub API guide for search, uploads and chat.
If something goes wrong
| You see | What it means |
|---|---|
| Generate key stays greyed out | The widget needs at least one collection under Knowledge base, a Name and a Greeting, plus a valid allowed origin starting https:// with no path. |
| This chat has not been given any knowledge to answer from yet. | The key has no collection to answer from, or the collection was deleted. Use Edit widget on the key and pick one. |
| Invalid embed key or origin not allowed. (where the chat should be) | The page address is not on the key's allowed origins, or the key was revoked. Generate a key for the exact address the page runs on. |
| A document shows Stored or Failed | Stored means it uploaded but cannot be read for answers (see What you can upload). Failed means reading it went wrong; delete it and upload it again, or try a different format. |
Reference
Building a deeper integration? See the developer guide for endpoints and events: