Agento Intelligence V1.10.7

Search and answers

  • Editors’ pinned search terms from Relevanssi now steer the assistant’s search, so a pinned page comes first for that term. Pins are used for ranking only and are never repeated in answers.
  • New “Sync search pins” button on the settings page, and a wp agento-intel sync-pins command, to push pin changes in seconds without a full reindex. Pins removed in WordPress are cleared too.
  • Two-letter searches such as IT, AI and HR now work. The minimum question length drops from three characters to two, and the on-screen message follows the setting.
  • Search results now show a content type label, the relevant date (added for documents, updated for everything else), the parent title for child tasks and projects, and a photo and job title for people, replacing the plain excerpt line.
  • Duplicate search results that share the same URL are removed.
  • The search shortcode accepts limit="…" to control how many results are returned, up to 50.
  • The RAG shortcode has a new limit parameter to set the number of sources per answer (default 5, maximum 15). The shortcode help now clarifies that type, category and tag take term slugs.
  • Thinking text accepts several pipe-separated phrases and picks one at random each time.
  • Answers are cached against the custom prompt, so editing the prompt no longer serves answers generated under the old one.

Reliable links in answers

  • Internal links that the model invents no longer appear as links. Any link to this site that isn’t a cited source, a real post or page, an existing uploaded file or a working URL is shown as plain text, keeping the label.
  • Links to uploaded files are checked on disk rather than over HTTP, so files behind access restrictions are not wrongly dropped.
  • Link checks forward the visitor’s login and ignore redirects, so private intranet pages verify correctly and WordPress’s slug guessing can’t let a fake path through.
  • Results are cached per user (a day for valid links, an hour for invalid) to limit repeat requests. External links are untouched, dropped links are logged, and the whole check can be turned off with the agento_intel_verify_internal_links filter.

New answer layout

  • Sources now sit in a right-hand column alongside the answer, collapsing to a single column on small screens.
  • The breathing search icon is replaced by the bot name and three bouncing dots, with a reduced-motion fallback.
  • Feedback controls appear directly after the answer in both branded and plain modes, with the disclaimer shown above them.
  • Name pills are replaced with plain bold labels for the user and the assistant.
  • Answers restored from history now look the same as live answers.

Feedback

  • Feedback is now stored in the site’s own database and reviewed under Tools → Agento Intel Feedback, with filtering, counts, CSV export, per-row delete and a details view (shared with the query logs page).
  • New AI summary of recent feedback, available to administrators only. Feedback is automatically purged after a year (filterable).
  • Thumbs up or down is saved as soon as it is clicked, so ratings are no longer lost when someone leaves without adding a comment. Changing your mind or adding a comment afterwards updates the same record instead of creating a duplicate, and the rating is restored if the answer is reloaded in the same session.
  • Sending a comment with no rating, or an empty comment, now explains what is missing rather than doing nothing.
  • Feedback edits are limited to the session that created them and to 24 hours.
  • The feedback panel closes itself after a comment is sent, stays open while you are still typing, and no longer gets stuck open with reduced motion or in a background tab.
  • Feedback buttons follow the page text colour, and the outlined submit button also applies when the theme is on auto and the system prefers dark.

More content in the index

  • Staff directory: a new “Index staff directory” setting adds people and teams to the index, with type="person" and type="team" available to the shortcodes. Profiles carry job title, grade, team, line manager, location, contact details, working pattern, skills, tags and bio. Teams carry description, lead, sub-teams and members. Profiles and teams update automatically when they are saved or deleted.
  • Published training courses and projects are now indexed, each with a settings toggle, live updates on save, trash and restore handling, and their own type filters in search, ask and WP-CLI.
  • WP-CLI reindex accepts person and team types and sends items in batches.

Better indexing

  • ACF values in indexed content now read properly: dates and times use the site’s formats, choice fields show their labels, and users, posts and terms show names rather than IDs.
  • “Expand” shortcode sections are flattened before indexing, so the toggle heading is indexed as content and shortcode markup is kept out.
  • Hidden elements are stripped from content before indexing.
  • Chunk size and chunk overlap are now settings, with sensible limits so passages are not cut short, plus matching filters.
  • Short trailing fragments are folded into the previous chunk instead of being indexed as near-empty chunks.
  • Indexing progress now counts every item handled, so runs no longer stall short of 100% when items have nothing to index, and the finished message says how many were skipped.
  • Saving a task reindexes it correctly again.

Content safety and security

  • Content safety screen: questions are checked before they reach the answering model. A new Content safety settings tab lets you enable the screen, choose blocked categories, and set a default refusal message and per-category messages (for example, wellbeing signposting for self-harm).
  • Content safety can also screen generated answers, using the same blocked categories and replies, and logs answer blocks separately.
  • Scan content for prompt injection: a new admin button scans indexed content for instructions aimed at the assistant, reports flagged items with edit links and snippets, and quarantines them.
  • Prompt-injection and persona detection is tighter and wider, catching role changes, named personas, accents, voice and style imitation, and poem or rhyme formatting requests, while ignoring innocent phrasing such as “how do I act as a fire marshal”.
  • New “This content discusses AI” tick box in the Agento Intelligence box on the edit screen (administrators only), so guidance about AI or prompts is indexed instead of quarantined.
  • Items allowed by an administrator are shown as a separate list in the content scan results, the WP-CLI scan output and the flagged-content email.
  • Password-protected posts are no longer indexed. Previously, an editor who had unlocked a page could cause it to be indexed for everyone.
  • Hidden users are skipped when indexing profiles.

Tools for editors

  • New excerpt generator: a Generate button in the classic editor Excerpt box writes a concise summary of the title and content, up to 50 words, for the editor to adjust before saving.
  • New tag audit: usage per tag, last used, authors and suggested merge groups by name similarity, available as a CSV download in settings and as a WP-CLI command.
  • The style checker now splits long content into batches, tells you up front when it will do so, and shows batch progress per rule.
  • The OCR tool has been shelved as we concentrate on language models rather than vision.

Settings and status

  • The status panel now names the models answering questions and warns when a model is missing from the server, instead of only showing Up or Down.
  • The status panel also reports the content safety state and per-type counts for users and teams.
  • Dashboard totals include training and projects, and show whether safety covers questions only or questions and answers.
  • Settings tabs renamed for clarity: Authentication is now Security, and Host URL is now Agento Intelligence Server.