Short notes on what we ship for customers. Newest first, grouped by month.
August 2026
23 August — Performance Style for text vs voice
Added Performance Style as separate text and voice prompt components, so the same agent can be concise on the phone and fuller in chat. Documented in Help under prompt components.
22 August — GPT-Realtime 2.1 for Voice Mode
Upgraded the realtime voice model to GPT-Realtime 2.1 so spoken conversations stay more natural under load. Voice still uses the same agents and knowledge as text.
22 August — Better chunks for URL curated answers
Improved pinned-URL curated answers so the agent reads the most relevant passages on that page, not just the first chunks. Intent matching is unchanged.
21 August — Staffed hours on the Human Handoff desk
Set weekday hours, a timezone, and closed dates so Human Handoff follows the clock. Typical setup is Human-Supported while the desk is open, and AI Only overnight, on weekends, and on holidays. Admins can override the live mode until the next scheduled change without rewriting the week.
21 August — Auto-assign waiting Human Handoff enquiries
On the Human Handoff screen, turn Auto-assign on and mark yourself Available. In Human Only every new chat is queued; in Human-Supported only escalations are. The oldest waiting conversation is claimed by the next idle agent, so nobody has to pick from a pile of cards.
20 August — Assistive answers after Human Handoff takeover
After you claim a conversation on Human Handoff, generate the same answer the customer would have seen, edit the markdown, and send it as a human reply. Citations stay in the admin pane. Generated drafts count toward usage; replies you type yourself do not.
20 August — Conversation tags on Human Handoff
Tag a live conversation from the Human Handoff panel (suggested tags from the service palette, or a custom label). Tags sit on the whole conversation, show on the queue cards, and stay on Insights after the session ends. Only the claimed agent can add or remove them.
20 August — Agent role for the Human Handoff desk
Give live-support staff an Agent login. They land on Human Handoff, work conversations and Insights, and can set Available or Away. They cannot change modes, hours, auto-assign, or other configuration.
20 August — Human replies stay in Insights
Recorded human-rep messages in Insights and stopped takeover from cutting off an in-flight AI stream, so the full thread is reviewable after a handoff.
20 August — Clearer markdown and inline images in answers
Improved markdown rendering and inline images so lists, emphasis, and photos in answers read more cleanly in the widget.
August — Visitor profiles on Human Handoff
When Visitor Profiles is on, a claimed conversation shows the captured attributes and a short summary above the transcript. Queue cards can show a visitor label so the rep does not start from scratch.
18 August — Show more / Show less on long facet lists
Added Show more / Show less on descriptor-driven facet lists so long filter panels stay usable. Also stopped the page jumping when visitors click Ask, and added thousand separators on counts.
16 August — Force one required tool call
Allow “Force call first” across multiple functions so an agent must run one tool before answering. Useful when lookup or escalation should never be skipped.
16 August — Hardened reCAPTCHA on chat and Ask
Fail open if Google reCAPTCHA is down, and dedupe widget tokens, so real visitors are not blocked by an outage while abuse protection stays on when Google is healthy.
15 August — Conversation traces in Insights
Persisted conversation trace logs and surfaced them in Insights so you can see how a live session was routed and retrieved, not only in Admin Chat.
15 August — Locale-aware search chrome and natural questions
Added a search-chrome i18n catalog driven by ui.locale, and allowed semicolons in questions so natural-language queries are not rejected. Short catalog replies can use each scope’s “view all” label.
15 August — Polite per-host crawl delays
Added per-host crawl delays and respect a capped robots.txt Crawl-delay, so large sites are indexed without hammering a single host.
14 August — Cleaner bookmarkable search URLs
Renamed the search-scope URL parameter to scope and restored “View all search results” after that change. Bookmarkable links stay stable for sharing and campaigns.
13 August — Tighter citations, welcome icons, and search links
Fixed double citations and agent-label placement, let welcome messages use icons, and omitted default parameters from bookmarkable search links. Empty-query summaries no longer fall back to a generic description.
13 August — More reliable Human Handoff connections
Fixed a disconnect that dropped operators mid-conversation so takeover stays on the same thread.
11 August — Smoother Human Handoff UX
Fixed handoff bugs and tightened the operator experience so claiming and replying feels less fragile under load.
10 August — In-flight search feedback and nested chunks
Added in-flight UI while search is running, batched search sub-requests, and nested chunk indexing so large sites return faster, more precise hits. Also fixed list formatting in answers.
10 August — Human Handoff stays up on long shifts
Reduced handoff load, fixed SSE watcher expiry, and plugged a memory leak so the desk can stay open without going silent.
9 August — Search on large sites and pagination restore
Fixed KNN depth on large university sites, improved hybrid search reliability, and restored inline-search pagination after reload.
8 August — Guided search View-all and cleaner citations
Fixed guided-search View-all links and tightened the search/citation UI so visitors land on the filtered results they were shown.
8 August — Smarter crawls (breadth-first, trap avoider, chunk packing)
Crawls now prefer breadth-first coverage, skip empty/boilerplate “duplicates”, cap noisy URL parameters, and pack related chunks. Index quality improves on large, parameter-heavy sites.
7 August — Sharper images from ingest
Improved how representative images are extracted and served (with CDN caching), so result cards and answers show the right photo more often.
6 August — Saved search results
Visitors can save Site Search results as they browse, then review the shortlist or email it to themselves. Admins turn it on under Search Configuration (label, max items, retention, optional email).
6 August — Search Guide for large catalogues
When a broad question matches hundreds of catalogue items, a Search Guide agent groups results, asks narrowing questions, or compares a short list, then links through to live filtered Site Search.
6 August — Validated fielded catalog queries
Agents keep the open goal in semantic search, then build a structured plan (scope, filters, ranges) that is validated against the live index so constraints such as campus, mode, or duration are not dropped as keywords.
5 August — Conversational Intelligence clustering fixes
Fixed clustering and Conversational Intelligence bugs so Categories & Themes stay stable across months.
3 August — Dates from JSON-LD and faster widget start
Indexed date metadata from JSON-LD, showed pagination even before a query, and stopped waiting for full page load before the widget initialises.
2 August — Welcome-chip icons and expanded hover widget
Welcome suggestion chips can show icons. Voice Mode now stops when the hover widget closes, and the widget can open already expanded.
1 August — Send questions into Site Search
Visitors can send a question from the widget into the search UI so Search and Ask stay one journey.
July 2026
31 July — No white flash in embedded search
Primed the iframe colour-scheme so inline search no longer flashes white on dark host pages.
30 July — Time-aware search for agents
Agents can apply a date range (and timezone) implied by questions such as “what’s on this weekend”, and bias toward recency for questions like “who is the CEO”, so answers are not drawn from stale pages.
30 July — Embed attributes for agent override and icon suppression
New data attributes let a page override the agent and hide the chat icon (for example on a search results page). Also tightened Ask viewport fit and catalog narrowing.
29 July — Event dates, badges, and guided search
Fixed missing date badges, date-range sorting on Events, skip-link junk in summaries, and guided-search function issues.
28 July — Result-card customiser save and clean field text
Fixed save in the result-card customiser and decoded HTML entities in field values so summaries no longer show & and friends.
27 July — Date-range search on result tabs
Visitors can search and sort by date ranges on scopes such as Events, with more reliable curated-answer matching on those journeys.
26 July — Clickable result images and $ badges
Search-result images now open the result page. Result-card badges can show a $ icon. Human Handoff gained further hardening (including SSE credentials).
25 July — reCAPTCHA on public chat and Ask
Enabled Google reCAPTCHA v3 on public chat and Ask so automated abuse is filtered without extra friction for real visitors.
23 July — Inline Ask welcome and scroll restore
Improved the inline Ask welcome and restored scroll position so switching Search ↔ Ask does not lose your place.
22 July — More date presets and honest search input
Added more date presets, aligned the live query with what is in the search box (not a stale URL), and fixed guided/embedded search and voice-in-forced-search analytics.
21 July — Choose metadata vs query-biased descriptions
New admin control: result descriptions can come from page metadata or a query-biased snippet. Also fixed inline images on Frontline and snippet generation.
18 July — “+N more” on result-card chips
Overflow chips became a real “+N more” link so visitors can see the rest of a multi-value field.
17 July — Reset conversation on Search, spelling, richer cards
Added a conversation reset on the search UI, hid the mic on Ask when Voice Mode is off, fixed spelling suggestions, and expanded result-card customisation.
16 July — Ask a question from your site search box
Visitors can type a question in a non-Airgentic (or header) search box and land in Ask, not only keyword results.
15 July — Stale and redirected pages leave the index
Removed 301 targets and stale pages from the index so search and answers stop citing dead URLs.
13 July — Fields from JSON with HTML fallback; facet fixes
Search Fields can be defined in JSON with HTML metadata as fallback. Facet and admin-chat auth bugs that blocked filters are fixed.
12 July — Human Handoff autoscroll and live agent indicator
Autoscroll when a rep locks or releases a conversation, show when the agent is responding, and avoid duplicate messages.
10 July — Search Insights and URL-length ranking for search
Search Insights is live (queries, clicks, no websocket required). The URL-length slider now applies to search ranking as well as RAG.
9 July — Scopes from any field; tighter curated matching
Scopes can be based on any indexed field, not only page type. Curated answers embed the question only and search separately, so the wrong override is less likely to fire.
8 July — Multi-value facets and mobile filter accordion
Facets can be multi-valued. Filters collapse into an accordion on mobile. “Force search” now applies on every conversation turn, not only the first.
7 July — Clean titles (entities and stray markup)
Decoded HTML entities and stripped leftover tags from result titles so cards read as the page intended.
6 July — Full titles and greyed-out empty scopes
Titles are no longer truncated. Scopes with zero hits grey out. Curated-answer selection is more accurate.
4 July — Voice Mode: echo, silence, and telephony
Enhanced Voice Mode with echo cancellation (including Twilio), faster RAG on transcript arrival, and fixes for permanent silence and welcome-message volume dips.
3 July — Category-based document age penalties
Admins can penalise older documents by category (news vs policies), so “who is the CEO” prefers recent pages without burying evergreen content.
2 July — Frontline can act as the expert
When there is only a Frontline agent, it can act as the expert instead of failing to delegate. Same setting as “Frontline as expert” later in admin.
June 2026
29 June — Date filters with time, and hide chat on search pages
Better date indexing (including a time field) and date-filter configuration. Chat icon can hide on search pages. Mobile search margins and AI summary are fixed.
28 June — Images in answers and thinking ellipsis
Agents can place inline images from cited pages, with an ellipsis while thinking. Result cards support category mapping; date facets and style flicker are fixed.
26 June — Microdata indexing and recency for undated pages
Indexed HTML microdata and stopped treating undated documents as brand new. PDF hangs and field-editor issues in the search configurator are fixed.
25 June — Style editor and AI prompt-change review
New style editor, a red unsaved-changes bar, and an AI review of prompt diffs before you save.
24 June — Follow-up support and a more prominent Ask tab
Enabled customer follow-ups in AI support and made the Ask tab more obvious. Images and welcome height tweaks for production sites.
23 June — Hierarchical filters, recency bias, multiple citations
Hierarchical filters can be built from several fields. Recency biasing is stronger, and answers can cite multiple passages (including from document bodies). Search UI design refresh.
21 June — New CSS architecture for search and widget
Rebuilt style tokens so brand colours, dark/light search, and widget chrome stay consistent.
19 June — GTM dataLayer, exclude URL patterns, HTTP functions
Push chatbot events to window.dataLayer for GTM/GA4. Crawler exclude URL patterns. Agents can call HTTP via curl-style functions. Optional Frontline-as-expert setting.
18 June — Guided search links from the agent
Agents can emit search links (scope + filters) so a visitor opens the live results the agent planned. CORS for the new search UI is fixed.
17 June — Descriptor field definitions
Search descriptors load field definitions so filters and cards stay aligned with the live index.
15 June — PDF/Doc icons and categories as filters
File-type icons on results, categories usable as filters, and mobile search formatting fixes.
14 June — Site Search polish: cards, facets, Ask-from-link
Many Site Search improvements: card edit/preview, facet correctness, hide welcome/title when arriving via a direct question, and smaller filter labels.
12 June — Voice Mode keeps accent and {{title}}
Voice Mode maintains accent instructions and substitutes {{title}} correctly. Transcription and fallback-to-Frontline when delegation fails are more reliable.
8 June — WhatsApp integration
Visitors can reach the same agents on WhatsApp, with the conversation logged like web chat.
5 June — Hide widget when a service is disabled
The embed hides itself when the service is disabled. Account owners can delete a service from the console.
2 June — Voice Mode on GPT-Realtime-2
Many Voice Mode fixes on gpt-realtime-2: speech bubble, spacing between turns, and quieter rooms. Search queries are no longer mixed into conversational analytics.
1 June — Faster scope tabs (no re-summary)
Switching search tabs no longer reloads the AI summary; tabs show they are switching. Example-question links work again.
May 2026
31 May — Inline search page and agentic retrieval
Added search_inline.html for a dedicated results page, plus agentic retrieval so the assistant can plan follow-up searches. Example-question borders are darker for contrast.
30 May — Stronger result summaries and dark mode
Enhanced search-result summaries, improved image fetching, and fixed dark-mode design. Offline search no longer fails closed.
29 May — Any admin can start a crawl
Removed the superuser-only gate on starting crawls so customer admins can refresh their index.
28 May — Document upload for all admins; description metadata
Non-superusers can upload documents. HTML processing now uses description metadata. Larger documents get longer processing timeouts.
27 May — Customers can add services
Any customer account can add a service, not only Airgentic staff.
26 May — Cleaner citations and Linear RAG ranking
Internal .answer URLs stay out of citations. Optional Linear Retriever with RAG, category boosts, and URL-length ranking improve what the agent reads.
24 May — Contextual welcome and faster hover widgets
Contextual welcome messages work again. Hover widgets bootstrap faster (hello/customisations over REST) with reconnect logic if the stream drops.
22 May — Conversational Intelligence for every customer
Clustering and themes run for all customers. Partial conversations are recorded so abandoned chats still appear in Insights.
20 May — Voice Mode noise handling (including iPhone)
Adaptive noise detection and quiet-room fixes so Voice Mode works in noisy spaces and on iPhone.
19 May — Query simulator and RAG ranking in Search
Admin query simulator for testing search, synonym save fixes, and optional RAG ranking for search itself. Curated-answer workflow updates.
17 May — Autocomplete opens document titles
Clicking a document title in autocomplete goes to that page. Native autocomplete follows the search theme and works on Wix via an extra data attribute.
16 May — Autocomplete on your existing search box
Airgentic suggestions on native search inputs (data-autocomplete-input-id and related attributes), not only inside the Airgentic overlay.
15 May — Dark theme on Service Hub and mobile
Dark theme applies to Service Hub; mobile and dark-theme contrast improved.
14 May — Search UI after launch: Ask height, filters, summaries
Post-launch polish: taller AI answer, scope shown on Ask, less confusing filter editing, body summaries when description is too long, and auto-categorisation toggle.
13 May — Honour noindex; cleaner titles
Crawler respects noindex. Category names no longer prefix search result titles.
11 May — Faceted Site Search in admin
Teams can build faceted search in the admin console (metadata, URL structure, dates, numbers, categories, hierarchies) without a custom implementation per use case.
11 May — Search journeys, scopes, and filters
Different journeys per content type (courses, people, news, policies, services) with filter controls such as hierarchy navigation, ranges, checkboxes, multi-select, and toggles.
11 May — Curated results and result cards
Admins can promote important results and customise how hits are displayed (result cards, badges, layouts per scope).
11 May — Search or Ask in one interface
Visitors can browse traditional results or ask the integrated AI agent from a single search UI.
8 May — Agent date search, timezone, and OIDC tokens
Agents can run advanced date searches with the visitor’s timezone. Full conversation threads record correctly; stale OIDC tokens no longer break secure embeds.
7 May — Overlay search DOM compatible with existing site CSS
New overlay DOM/styles so existing site stylesheets (including Facebook-style layouts) apply with fewer overrides.
6 May — Context-sensitive Help in admin
Admin screens link to the matching Help article. (Curated Results / filters / field boosts from this week sit in the published 11 May launch.)
April 2026
30 April — Sleeker chat and search theme
A new themes system ships with an April 2026 look: cleaner overlay and widget chrome so chat and site search sit more naturally on customer sites.
30 April — Smoother voice conversations
Voice mode is more reliable, with a clearer transcript prompt and fixes so spoken answers stay in order and interruptions work as expected.
29 April — Facet filters for site search
Site search gains a new facet-filter system so visitors can narrow results by the fields that matter for your catalogue, courses, or services.
27 April — Faster first paint for chat and search
Airgentic initialises more quickly and takes over the site search box sooner, so the widget feels ready on the first interaction.
26 April — Enterprise search for existing search platforms
Enterprise search ships: search descriptors, category navigation, and compatibility with existing search platforms so organisations can run a governed site-search experience from the admin console.
25 April — Smarter crawls and richer result cards
Incremental crawling is faster and more reliable, crawl-trap avoidance is tighter, and result cards can use favicons as logos with a better social preview.
21 April — Monthly insights digest emails
Conversational Intelligence can email a richer monthly digest — with working links back into Insights — so teams see spikes, gaps, and emerging themes without opening the console first.
19 April — Documents Ingestion API
A documented API (with an API key) lets other systems upload files for indexing or custom processors — the programmatic counterpart to Upload Documents.
15 April — Human Handoff for every customer
Human Handoff is no longer gated: any customer can monitor live conversations and take over when someone needs a person.
13 April — Conversational Intelligence (limited beta)
Conversational Intelligence turns conversation volume into themes and categories overnight: LLM-named themes and clusters, scheduled analysis, and a customer guide so teams can act on demand — not just ticket counts.
12 April — Categories & Themes insights
Insights now has Categories & Themes: demand and evolution charts, overlap, sentiment filters, drill-through to conversations, and clearer navigation between governed categories and discovered themes.
8 April — Phone support options for all users
Telephony settings are visible to all customers, not only superusers, so phone-based AI support can be turned on from the same service configuration as chat.
6 April — Analytics API can include user IDs
The Analytics API can store and fetch a user identifier, so exported conversations can be joined to your own CRM or identity records where you choose to send one.
2 April — SharePoint web part: choose the Service Hub
A new SharePoint web part (two-part package) lets you set the Service Hub location, so intranet pages open the right Airgentic experience.
March 2026
23 March — Force a function when you need it
Agents can be told to force a function call, so critical actions (lookup, escalate, submit) run even when the model would otherwise skip them.
22 March — User-segmented prompts
Prompts can vary by visitor country, region, or city, so the same agent gives location-appropriate guidance from the first message.
14 March — Custom chat launcher on your pages
The widget supports data-chat-icon-button-id and data-chat-icon-visible, so you can open chat from your own buttons and hide the default icon when you want a native page control.
14 March — Clearer customer onboarding
A new onboarding path (prep vs live session) and go-live checklist help teams get from first login to a production service with less back-and-forth.
12 March — Semantic clustering and Trends
Semantic clustering (phases 1–4) and a redesigned Trends experience — plus completion emails — give an early view of what customers are asking, before the April Categories & Themes polish.
11 March — Single sign-on without a second login
SSO can silently fail for unauthorised visitors (no extra login wall for the public), and authorised users no longer need a separate Airgentic login step. Help reflects the new mode.
11 March — Clearer, more natural phone voice
Voice on the phone uses the newer realtime speech model so spoken conversations sound more natural and stay in sync with the same agents and knowledge as chat.
9 March — Upload documents and crawl file shares
Admins can upload documents (public or secure), crawl a file system, and send mail from functions. Citation links to secure files authenticate correctly.
8 March — SharePoint web part and voice privacy
A new SPFx web part embeds Airgentic in SharePoint. Audio conversations are no longer recorded by default.
4 March — Customer functions (AirScript)
Customers can write their own AirScript functions — search, email, HTTP, escalation — so agents take actions, not only answer questions. Functions are live with customer docs.
1 March — Phone channel (Twilio)
Airgentic can take customer calls on a phone line, using the same agents, prompts, and knowledge as the website — with analytics in Twilio mode.
February 2026
22 February — AI question input fields
Pages can host question input fields that send the visitor straight into Airgentic with their question already typed — turning a static form or banner into a conversation.
22 February — Agents start in test; Help Centre is public
New agents start in test mode so they cannot go live by accident. The admin console has a search box, and the public Help Centre (help.airgentic.com) is live.
15 February — Prompt review assistant and modular prompts
Every prompt change can be reviewed by an AI assistant (summary, risks, suggestions). Prompts can include other prompts, so tone, guardrails, and retrieval stay consistent across agents.
14 February — Human Handoff in the admin menu
Human Handoff is exposed as a first-class screen (live monitor, take over, notifications) and the subscription menu item is hidden so the path to live support is obvious.
12 February — Inline citation links
Answers show inline citation links (including multiple sources), so visitors can open the exact page or document the agent used.
10 February — Secure services: auth, UI, and images
A new secure authentication and authorisation model, secure UI, and secure PDF page images so intranet and authenticated knowledge stays behind your identity provider.
8 February — Prompt version control
Administrators can see a prompt at any point in time, compare revisions, and restore an earlier version — so small prompt edits are safe to iterate.
8 February — Citations when a specialist agent answers
When an expert agent answers, the visitor still sees a citation, not a bare reply.
7 February — Delete indexed URLs
Admins can delete specific URLs from the index when a page should no longer be searchable or cited.
6 February — Trace log for every enquiry
Full trace logging shows search results, prompts, and function calls for each enquiry in Admin Chat, so teams can see why an answer was given.
4 February — Voice mode reliability
Voice mode scrolling, microphone flicker, and function-call errors are cleaned up so spoken sessions stay usable on the widget.
January 2026
22 January — New citation system
A rebuilt citation experience (with scroll-to-source) makes it obvious which page or passage grounded the answer.
20 January — Analytics API and custom chat placeholder
Conversation analytics can be pulled via API. The chat input placeholder is configurable per service.
17 January — Human Handoff
Staff can watch live conversations, get alerts (thumbs down, negative sentiment, pause expiry), claim a thread with locking, toggle the AI assistant, and reply as a human — including Human Only mode.
13 January — Account subscriptions customers can manage
Customers can cancel or change a subscription, see entitlements on the subscription screen, and understand usage against allowances. Buying a paid plan cancels the overlapping trial.
12 January — Usage meters and add-on packs
Permanent subscriptions and add-on packs, clearer product names, and usage notifications only when a meter exceeds 100% — so customers see what they have bought and when they are over.
8 January — AI Visibility: richer page analysis
AI Visibility adds content-enhancement actions, long-sentence and social-metadata analysis, robots as its own check, and better canonical/performance reports — plus live overage alerts.
8 January — Insights filter by website
The insights dashboard can filter conversational threads by website, useful when one account serves more than one property.
7 January — Crawl status and usage at account level
Data sync shows crawl status. Usage and entitlements are account-wide so seats and meters are shared across services on the same subscription.
5 January — Content gap analysis
A content gap analysis (with a clearer report) shows where customer questions are not backed by indexed pages.
5 January — Customer admins can delete an account
Account owners can delete their own account from a dedicated management page, with a short feedback survey.
2 January — Developer & API documentation
Public developer and API docs land so integrators can work from a published reference rather than private notes.
2 January — More reliable website crawler
A requests-style crawler improves how public sites are fetched for search and answers.