Before you use the Agento Intelligence plugin, you must set up security with Agento Digital first. Agento Digital gives you an API key and the address of your Agento Intelligence server. The server only accepts requests from web servers it has been told about, so tell Agento Digital where your intranet is hosted.
Agento Intelligence plugin
Install and activate the Agento Intelligence plugin. You’ll find the Agento Intelligence settings on the Options menu. There are 7 tabs: Settings, Personalisation, Writing tools, Content safety, System status, Index and Security.
Set the plugin up in this order:
- Enter your API key and server address on the Security tab.
- Choose what to index on the Settings tab.
- Build the index on the System status tab.
- Add the assistant to a page with a shortcode.
Security
API key
Enter the API key supplied by Agento Digital. Keep your key safe and do not share it.
Agento Intelligence Server
Enter the server address supplied by Agento Digital.
Access mode
Set the Access mode depending on your intranet setup:
- Logged in users only: staff must be logged in to the intranet to ask questions. Use this if staff log in, including through single sign-on.
- Public access, not logged in: staff do not log in, and the intranet can be reached from the internet. Questions are only accepted with the public token, and each address is limited to a number of questions a minute.
- Private access, not logged in: staff do not log in, and the intranet can only be reached from your own network. Questions are accepted from anyone who can reach the intranet, with a limit on the number of questions a minute.
Public token
This box appears when the Access mode is Public access. Select Generate token to create a token, then save. The assistant sends the token with every question, and the plugin refuses questions without it. Generate a new token at any time to replace the old one.
Settings
Shortcodes
The top of the Settings tab lists the shortcodes that put Agento Intelligence on a page.
[agento_intel_rag]adds the assistant, where staff ask a question and get an answer with links to the pages the answer came from.[agento_intel_search]adds AI search, which lists the best matching pages and people without writing an answer.
Both shortcodes accept optional settings to narrow what they look at:
type: task, page, training, project, media, person or team. For example[agento_intel_rag type="person"]makes an assistant for the staff directory. Separate several types with commas.categoryandtag: the slug of a category or tag, such ascategory="human-resources".limit: for the assistant, how many pages an answer can draw on. The default is 5 and the most is 15. A higher number suits staff directory pages. For AI search, how many results to list. The default is 10 and the most is 50.
The image-to-text shortcode from earlier versions has been retired and produces nothing.
Index tasks
Choose whether to index tasks and guides.
Index pages
Choose whether to index pages. You must select which pages to include under Pages to include. Descendants of selected pages are also included.
Index training courses
Choose whether to index published training courses, with their target audience, prerequisites, learning objectives, duration and booking details. This option needs the Training module.
Index projects
Choose whether to index published projects, with their status, dates and team members. This option needs the Projects module.
Index documents
Choose whether to index documents in the Media Library, such as PDF and Office files. Note that this feature only indexes the title, description and file name of documents; it does not index the contents of documents.
Index staff directory
Choose whether to index staff profiles and teams, so that staff can ask questions such as “Who is the head of finance?” or “Who are the first aiders in Bristol?”.
- Staff profiles include name, job title, grade, team, line manager, location, contact details, working pattern, skills, tags, biography and any extra profile fields.
- Teams include their description, team lead, sub-teams and list of members.
- Only details that the staff directory already shows to logged-in staff are indexed. Hidden users are left out.
- To make an assistant or a search that only looks at people or teams, use
type="person"ortype="team"in the shortcode.
Collect feedback
Choose whether to show thumbs up and thumbs down buttons under each answer, with space for a comment. Feedback is anonymous and is stored only in your intranet’s own database. It is not sent to Agento Digital. Review feedback under Tools, Agento Intel Feedback. Feedback is kept for 365 days.
On the feedback report, Analyse negative feedback sends the questions and answers that were marked as unhelpful to your Agento Intelligence server, which suggests how to improve your content. Nothing is sent unless you select that button.
Feedback style
This option appears when Collect feedback is on. Choose Plain or Branded for the look of the feedback buttons. Branded uses your intranet’s colours.
Track analytics
Send questions to Google Analytics or Matomo as events, so that you can see what staff ask alongside your other statistics.
Log questions
Turn on logging. Reports are available under Tools, Agento Intel Logs. This tracks questions and responses for the last 30 days. Personal details such as email addresses and phone numbers are removed from questions before they are saved. Security events, such as a refused question, are always recorded, even when logging is off.
Allow history
This setting shows up to 3 recent questions for a user, for quick access to commonly used guidance. Recent questions are remembered in the user’s own browser, not on the intranet.
Batch size
How many pieces of content are sent to the server at a time while indexing. The default is 10. Leave this as it is unless Agento Digital suggests a change.
Chunk size
Content is split into chunks for the index. Chunk size is the most characters in a chunk. The default is 800. Larger chunks give the AI more of each page in one passage, but keep the number at or below 1600. Changing the chunk size requires a full reindex.
Chunk overlap
The number of characters repeated between one chunk and the next, so that a sentence on the boundary is not lost. The default is 120. Changing the chunk overlap requires a full reindex.
Cache period
Remember questions asked by users and show the same answer without having to query the AI system again. Enter the number of hours to remember an answer for. Set to 0 for no cache. Answers are never cached for administrators, so that you always see a fresh answer when testing.
Personalisation
Thinking text
Enter the message to appear when the AI is ‘thinking’. Example: “Checking intranet content…”. To vary the message, enter several phrases separated by a pipe, and one is chosen at random each time. Example: “Checking intranet content…|Looking that up…”
Query examples
Add example questions to prompt users, separated by a pipe. They appear one after another as placeholder text in the question box. Example: “How do I claim travel expenses?|What pension schemes are available?”
Disclaimer text (optional)
Enter disclaimer text to appear after every answer. Example: “AI can make mistakes. Always check source material.”
Personalise AI chat
Turn on if you want to give the AI function a name. The person’s own name is then shown above their question and the assistant’s name above each answer.
AI assistant name
This box appears when Personalise AI chat is on. Enter the name to show above each answer. If you leave it empty, “AI Assistant” is used.
Change the system prompt
Turn on to add your own instructions for the assistant. A System prompt box appears. Use it to add additional AI prompts, for example, to celebrate Pride month or to encourage staff to complete the staff engagement survey.
Your instructions affect the style, tone and structure of answers. They cannot override the built-in rules, such as answering only from intranet content.
Search icon
The icon on the button that sends a question. Enter a Font Awesome icon name without the fa- prefix, such as magnifying-glass, paper-plane or wand-magic-sparkles. Leave blank to use the default arrow.
Search icon style
This option appears when you have entered a search icon. Choose the Font Awesome style for the icon, such as Solid. The style must be one that your intranet loads.
Writing tools
Writing tools help editors in the WordPress edit screen. They use the same AI server as the assistant.
AI rewriter
Turn on to add an AI rewriter button above the editor, next to Add Media. An editor chooses a rule, selects Rewrite, and sees their content rewritten to follow that rule. They can then select Copy, Replace editor content to swap the rewritten version into the page, or Go back. Nothing is changed unless the editor chooses to use the result, and the page is not saved until they select Update.
Rules
This option appears when AI rewriter is on. Add one or more rules. Each rule has a Title, shown in the list that editors choose from, and a Prompt, which is the instruction sent to the AI. Example title: “Plain English”. Example prompt: “Rewrite in plain English. Use short sentences and the active voice. Keep all facts, headings and links.”
Excerpt generator
Turn on to add a Generate button to the Excerpt box on the edit screen. The button sends the title and content to the AI and fills in a summary of no more than 50 words, which the editor can change before saving.
PDF to page
Turn on to add a PDF to page tool to the Media menu and to each PDF in the media library, for anyone who can create pages or tasks. It reads a PDF and creates a draft with the text as headings, paragraphs, lists and tables, with the document’s links and footnotes, its illustrations if the editor asks for them, and a summary as the excerpt. Contents pages, page numbers and headers and footers are left out. The wording is copied from the PDF, not rewritten, and the draft is not published until an editor has checked it. The converter must be installed on your AI server. See Convert a PDF to a page.
Content safety
Screen questions for harmful content
This option is on by default. Every question is checked by a separate safety model before the assistant answers. A question in a blocked category is refused, recorded under Tools, Agento Intel Logs, and never reaches the assistant. The check judges what the person is trying to do, not the words they use, so questions about your drugs and alcohol policy or harassment guidance are still answered.
Screen answers as well
Also check each answer before it is shown, using the same blocked categories. This catches unsafe material that reaches an answer through indexed content. It adds a fraction of a second to each answer.
Blocked categories
Questions that fall into a ticked category are refused. These categories are blocked unless you change them: violent crimes, non-violent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, hate, suicide and self-harm, and sexual content.
These categories are available but not blocked at first: defamation, specialised advice (financial, medical, legal), privacy, intellectual property and elections. Think before you block specialised advice or privacy, because staff ask genuine questions about pensions, occupational health and data protection.
Refusal message
The message shown in place of an answer when a question is blocked. Use plain text, or simple HTML with links.
Category-specific messages
Replace the refusal message for particular categories. Choose a Category and enter a Message. For example, for suicide and self-harm, give the contact details of your employee assistance programme and the Samaritans.
System status
The System status tab shows whether the service is working.
- Intelligence server: whether your intranet can reach the server.
- AI models: the model that writes answers and its backup, and whether the server has them.
- Content safety: whether screening is on, whether it covers questions only or questions and answers, and how many categories are blocked.
- Status: Healthy or Unhealthy.
Index
The Index tab shows what is in your index and has the buttons that build and maintain it.
Statistics: how much content is in your index, in total and for tasks, pages, training, projects, media, users and teams. The number of chunks is higher than the number of items, because long content is split into several chunks.
Reindex all
Builds the index from all the content you chose on the Settings tab. Do this when you first set up the plugin, and after you change what is indexed, the chunk size or the chunk overlap. The tab shows how much has been processed and an estimated time to finish. Keep the tab open until it says Done, because the reindex is run from your browser and stops if you close the tab or move to another page. Select Cancel to stop.
Purge index
Removes everything from your index. The assistant cannot answer until you reindex.
Scan content for prompt injection
Checks your content for text that tries to give the assistant instructions, such as “ignore your previous instructions”. Content like this is kept out of the index. The scan lists anything it finds, and anything an administrator has allowed.
Sync search pins
If you use Relevanssi Premium, editors can pin a page to a search term so that it comes first in search results. Sync search pins sends those pins to the assistant, so that AI search and answers favour the same pages. It takes a few seconds and does not reindex. Pins are also sent whenever a page is indexed.
Download tag audit (CSV)
Downloads a spreadsheet of tags that mean nearly the same thing, such as “annual leave” and “holiday”, so that you can merge them.
Keeping the index up to date
After the first reindex, the index looks after itself. Content is indexed again whenever it is saved, and removed from the index when it is unpublished or deleted. Staff profiles are indexed again when a profile is updated, and teams when a team is saved.
Administrators see an Agento Intelligence box on the edit screen, with the option This content discusses AI. Allow instruction-like text and index it. Content that appears to give the assistant instructions is normally kept out of the index. Tick this box for genuine guidance about AI or prompts, then save to index it.
News and events are not indexed.
What the plugin adds elsewhere
- Tools, Agento Intel Logs: the questions asked and the answers given, if logging is on. Kept for 30 days.
- Tools, Agento Intel Feedback: the thumbs up and down, with any comments, if feedback is on. Kept for a year.
- The edit screen: the AI rewriter button, the Generate button in the Excerpt box, and an Agento Intelligence box for administrators. See Generate an excerpt.
- Media, PDF to page: turns a PDF into a draft page or task. See Convert a PDF document to a page or task.
- Search pins: pages that editors pin to a search term in Relevanssi are favoured by the assistant too. See Promote items in search results.
The assistant does not replace the intranet search box. That is still Relevanssi, set up under Search options.
Command line tools
If you have shell access to the web server, the plugin adds commands to WP-CLI. They do the same as the buttons on the Index tab, and suit large intranets where a reindex takes a long time, because they do not need a browser tab kept open.
wp agento-intel reindex: rebuilds the index from the content chosen in Settings. Add--types=task,pageto do only some kinds of content,--dry-runto see what would be indexed without sending anything, and--yesto skip the confirmation.--per-pagesets how many items go in each batch; the default is 15.wp agento-intel purge: empties the index. Add--yesto skip the confirmation.wp agento-intel status: shows the plugin’s settings and whether the server can be reached.wp agento-intel stats: how much content is in the index. Add--format=jsonor--format=yamlfor machine-readable output.wp agento-intel ping: checks the connection to the server.wp agento-intel scan-content: looks for content that tries to give the assistant instructions. Nothing is sent to the server. Add--typesto narrow it.wp agento-intel sync-pins: sends editors’ Relevanssi search pins to the assistant without reindexing.wp agento-intel tag-audit: writes a spreadsheet of tags that mean nearly the same thing. Add--outputfor the file path,--taxonomyfor a taxonomy other than tags, and--similarity, from 0.5 to 0.99, for how alike two names must be to be grouped. The default is 0.9.