MxChat – AI Chatbot & Content Generation for WordPress

3.2.22 – September 20, 2026

• New: An opt-in Pinecone Document index with full-text search, so a question carrying a part number, SKU or code finds the record that holds it. Create index for me, Check index and Migrate move an existing index across with no re-embedding, and classic vector indexes are untouched.
• New: The chatbot knows the current date and time in your site’s timezone, so it can tell past events from upcoming ones. The line travels with each message and does not affect prompt caching; {current_date} and {current_datetime} also work as placeholders in the instructions.
• New: Knowledge base content can be restricted to any role on your site, including custom ones from User Role Editor, Members or a membership plugin. The built-in Subscribers and Above through Administrators Only choices behave exactly as before.
• New: Every post and page has a Knowledge base access setting in the MxChat box on its edit screen, so content without tags is no longer stuck as public. Where both apply, the stricter of the tag rule and the page setting wins.
• New: PDF, Word, text and Markdown files already in your media library can be imported into the knowledge base, with the text read from the file itself. The Media source now lists only those four types.
• New: Switch on Media under Advanced Custom Post Sync Settings and a supported file adds itself to the knowledge base on upload, updates its entry when replaced and is removed when deleted.
• New: The pre-chat lead form can be shown to signed-in visitors too, pre-filled from their profile, and the email address can be made optional so you can collect a name and consent alone. Leads captured either way appear in the Leads tab and the CSV export.
• New: The API can remove knowledge base entries by source URL, singly or up to 50 at a time, so a document deleted in SharePoint, Drive or your CMS can be taken out of the chatbot by your sync job. Deleting a URL that is not stored is reported as nothing removed, not an error.
• New: Lead capture submissions can be read over the API, with paging and a since filter, so leads can be synced into a CRM or spreadsheet automatically.
• New: Gemini 3.8 Flash, 3.7 Flash and 3.6 Flash join the model picker, with 3.8 Flash as the recommended Gemini default and 3.5 Flash-Lite as the lowest-cost choice. Sites already on 3.5 Flash keep it until they pick something else.
• New for developers: Add-ons can put a button in the chat toolbar with the mxchat_chat_toolbar_items filter, and the mxchat:toolbar-ready and mxchat:opened page events mean no more watching the page or running requests on every view. The WooCommerce cart button is the first to use it.
• New for developers: Add-ons can append to a chat transcript with save_session_message, shape what humans read with mxchat_transcript_user_message and mxchat_ai_conversation_history, and add sections to the Telegram and Slack handoff card with mxchat_handoff_card_sections.
• Security: Hardened how links are written into the chat window and how a visitor’s own earlier messages are restored after a page reload. Ordinary links, encoded addresses and file names with brackets look and behave exactly as before.
• Security: Hardened per-visitor rate limiting and its hourly reset. A visitor’s allowance is now counted against the address the request really came from, with Cloudflare handled automatically, and daily, weekly and monthly limits hold for the full period; other proxies can use the new mxchat_client_ip filter.
• Fixed: Restricting a knowledge entry to a custom role no longer leaves it readable by every signed-in visitor.
• Fixed: Choosing Media in the WordPress Content importer now lists your files instead of coming back empty on every site.
• Fixed: With Pinecone, the Knowledge Entries heading shows the vector count Pinecone reports instead of 50 for any index over 500 vectors, paging uses Previous and Next and continues where the last page ended, and a bot with its own empty namespace shows 0.
• Fixed: With Contextual Awareness on, sending a message on a page with a video or audio player no longer starts an invisible copy playing or downloads the media again. The page text is read from an inert copy, gathered once per page.
• Fixed: The function-calling loop no longer sends an invalid request when the per-turn tool budget is reached. Every tool call gets a result, the model answers from what it has, and a provider failure after a tool has run is logged even with developer mode off.
• Fixed: Words joined with underscores, such as file names, SKUs and placeholders, no longer lose their underscores and turn italic. A backslash writes an underscore, asterisk, tilde or backtick literally, and anything inside backticks is shown exactly as written.
• Improved: The Hybrid keyword boost now applies to Pinecone document indexes as well as the WordPress database, and the Transcripts and Testing panels label each match Vector, Keyword or Both.
• Improved: With AI Tools active, questions that need no tool are generated once instead of twice, so replies arrive sooner and ordinary turns use roughly half the output tokens. Those replies arrive as one message; the mxchat_fc_keep_first_hop_text filter restores the old behaviour.
• Improved: Gemini 3 series models are asked for a low thinking level on every chat reply, natively and through the OpenAI-compatible provider, so you stop paying for hidden reasoning tokens. A developer can change it with the mxchat_gemini_thinking_level filter.
• Improved: The API Access screen documents every endpoint in one place, with ready-to-run examples for removing a deleted document and syncing new leads.
• Improved: The plugin listing’s appearance FAQ now says which customization options are in the free plugin and which come with the Theme Customizer add-on.
• Changed: The plugin now declares WordPress 5.3 and PHP 7.4 as its minimum versions, matching what the code has required for some time. Every MxChat add-on already required PHP 7.4.

3.2.21 – September 6, 2026

• Fixed: Ending a Telegram live agent conversation, from MxChat or from Telegram itself, now retires its topic and returns the visitor to the AI. The visitor’s next request used to post into the closed topic, where agents could not reply.
• Fixed: The conversation history shared with Telegram agents now always arrives, sent in parts within Telegram’s message limit as readable text. A long conversation could previously drop the entire handoff card without a trace.
• Fixed: Slack agents who reply in a thread on a handoff message now reach the visitor. Those threaded replies were silently discarded in the default per-conversation mode.
• Fixed: With Chat Persistence on, Start new chat now leaves only your greeting, and a restored conversation keeps the greeting above it. A returning visitor was previously left with one old reply and no greeting.
• Fixed: The Open Telegram settings button on the webhook security notice now opens the Telegram settings tab instead of the onboarding screen.
• Improved: Telegram handoffs share the last 10 messages, labeled with the count, and any Telegram delivery failure is shown on the Telegram settings screen with Telegram’s own explanation. Failures used to leave no trace anywhere.
• Improved: When your site rejects Telegram’s incoming requests, the Telegram settings screen and a dashboard notice now say how many, when, why, and how to fix it. Sites behind Cloudflare with no webhook secret previously dropped every reply in silence.
• Improved: When a Slack message cannot be relayed to the visitor, such as an edit or a file upload, the agent now sees a short note in the channel explaining how to get it through. Dropped replies are logged when debug mode is on.
• Improved: If your shared Slack handoff channel is private, the Slack settings screen now warns that your Slack app needs the message.groups event subscription. Without it, handoffs post but agent replies never reach the visitor.
• Improved: On the Pro and Extensions page, the button for an add-on you have not installed is now Get the ZIP from My Account and opens your account Downloads page. It used to open the same page as the Documentation button.

3.2.20 – August 28, 2026

• New: The knowledge base can now populate and maintain your OpenAI Vector Store — every addition, edit and deletion is mirrored automatically, and an Import Existing Knowledge Base button (or the wp mxchat vectorstore-import command) uploads what you already have in resumable batches; role-restricted entries are never uploaded.
• New: Chat answers can no longer link to pages that do not exist — every finished answer, streaming included, is checked and links to your own site that would 404 are removed while the sentence stays intact.
• New: A Strip Unapproved Links setting separates link enforcement from Citation Links, so invented links are removed even with citations off; on for new installs, opt-in for existing sites.
• New: A third live-agent handoff destination, Webhook — send handoffs to your helpdesk, CRM or automation flow as a signed JSON POST, with its own availability schedule and delivery errors shown on the settings screen.
• New: The pre-chat email form can show a consent checkbox with your own label and Privacy Policy link, optionally required, enforced on the server — and each capture records whether it was ticked, when, and the exact label shown, visible in Leads, the CSV export and privacy exports.
• New: The chatbot can be placed with a Gutenberg block, or an Elementor widget on Elementor 3.5+ — inline or floating, with a bot picker when Multi-Bot is active, and a neutral placeholder in the editor instead of a live chat.
• New: A Document Upload import method takes .docx, .txt and .md files straight into the knowledge base; the file itself is never stored and re-uploading the same name replaces rather than duplicates.
• New: A Custom Retention (Days) field on Transcripts lets you delete transcripts after any number of days instead of only the weekly and monthly presets.
• New: Pinecone knowledge entries can now be edited in the admin — saving re-chunks and re-embeds in place, so fixing a typo no longer means delete and re-import.
• New: Each ACF field group gains a group-level toggle that includes or excludes every field in it at once, with a distinct half-way state when only some are included.
• New: Add-on AI tools now receive the model’s full tool-call arguments, so a tool declaring typed parameters gets them delivered instead of everything but the query being discarded.
• New: Three developer hooks for custom retrieval — mxchat_retrieval_query rewrites the lookup text without touching the transcript, mxchat_pinecone_query_body tunes Pinecone queries, and mxchat_session_ended fires on both deletion and retention expiry.
• New: The mxchat_final_response_text filter runs on the complete answer on every response path, streaming included, before it is shown and saved.
• Security: With Contextual Awareness on, page text captured from the visitor’s browser is now fenced as untrusted reference material, capped at 8,000 characters and required to belong to your own site, so instructions hidden in comments or reviews are far less likely to be followed.
• Fixed: ACF field toggles now identify each field precisely, so same-named fields in different groups no longer share one toggle or switch each other back off; existing exclusions are migrated on the safe side with a one-time notice.
• Fixed: All remaining Pinecone reads use the correct call shape and respect a configured namespace, so listings, chunked answers and already-indexed detection work on serverless indexes instead of silently returning nothing.
• Fixed: Long manually pasted entries each keep their own identity — previously a second long pasted document silently deleted the first one’s content, in Pinecone and the local database alike.
• Fixed: Shortening an entry below the chunking threshold now removes its old chunk records, so answers can no longer draw on stale text alongside the new version.
• Fixed: Knowledge entries whose address contains percent-encoded characters, such as non-English permalinks, no longer go missing or save a duplicate under a different identity.
• Fixed: Importing Persian, Hindi and other Indic-script content no longer splits words at invisible joiner characters; re-import content added earlier to repair it.
• Fixed: The pre-chat email form now scrolls when taller than the chat window, so the submit button stays reachable — especially on phones with the keyboard open.
• Fixed: Product cards from filtered WooCommerce searches display correctly after a refresh instead of showing raw markup, and history replay now recognizes every MxChat card type automatically.
• Fixed: Cards from an AI tool answer keep their position below the answer text when the chat is refreshed or reopened, matching the order the visitor originally saw.
• Fixed: A floating chatbot placed by shortcode or block on a page that also auto-displays one no longer renders two overlapping bubbles with duplicate element ids.
• Fixed: With citation links on, real published pages on your own site and web-search source links are no longer stripped from answers as unknown.
• Fixed: The Sources tab in Message Context now lists the documents behind tool-answered replies, and tells apart a search that found nothing from a reply with no recorded retrieval.
• Fixed: The “Translation loading for the mxchat domain was triggered too early” notice on WordPress 6.7 and newer is resolved.
• Fixed: A custom retention value set by hand or by code is no longer reset to zero when other transcript settings are saved.
• Fixed: Status pills on admin screens outside API Access now render with their intended styling instead of as plain text.
• Fixed: The chat message input has a real accessible name for screen readers, rather than relying on placeholder text that disappears on typing.
• Fixed: Sites that saved the AI Tools screen on earlier versions now receive the recommended usage guidance for the handoff tools; hints you wrote yourself are never changed.
• Improved: Screen reader support for the conversation is complete — a streamed answer is announced once when it finishes rather than continuously as it types, the waiting state and error notices are read out, and restored history stays silent.
• Improved: On phones and tablets the chat input no longer takes focus after each reply, so the keyboard stays closed while the visitor reads; desktop is unchanged and the mxchat_autofocus_after_reply filter forces either behavior.
• Improved: Saving a WooCommerce product no longer generates two knowledge-base embeddings for the same product, halving embedding API usage on product saves.
• Improved: The daily model health check now watches every model in use, including per-bot models from Multi-Bot, via the new mxchat_models_in_use filter.
• Improved: Replacements for retired AI models live in one shared mapping used across the plugin and its add-ons, and content generation is now rescued from any retired model.
• Improved: The Content Generation model picker derives from the central model catalog, so new models appear automatically and a saved model that is no longer offered stays selected and labeled.
• Improved: The Video Card Match Threshold is now a slider with a live value readout, matching the other threshold controls; its range and limits are unchanged.
• Improved: The admin Testing panel reports how many links the guard checked and which it removed, which was previously silent.
• Improved: On non-English sites, the chat widget’s default texts are saved in the site language on a fresh install.
• Improved: A Pinecone listing call that genuinely fails now logs that the fallback fired instead of masking it.

3.2.19 – August 15, 2026

• Fixed: Hebrew, Arabic and other right-to-left PDFs no longer import with reversed text, and a new WP-CLI repair command fixes entries imported before this update in place.
• Fixed: Sites still set to the retired Claude Opus 4.1, or to older retired Claude models, are automatically moved to the current Claude generation of the same tier and shown a notice.
• New: MxChat checks once a day that your selected AI model still exists and warns you on the AI Models screen and dashboard before a retired model can silence your chatbot.
• Improved: Claude API costs drop on most sites — Anthropic prompt caching now runs automatically, ahead of Anthropic’s September 1 Claude Sonnet 5 price change.
• Fixed: Content generation no longer fails with an Unsupported value error on GPT-5.4 Mini and GPT-5.4 Nano, and any model that rejects a reasoning setting is retried once without it.
• Improved: The xAI Grok lineup is current again — Grok 4.6 and the rest of the new generation are offered first, and existing selections keep working.
• New: Settings edits are protected — changed fields show an Unsaved marker, leaving mid-edit asks for confirmation, and edits are handed to the browser for saving if you leave anyway.
• Fixed: Imported content keeps its currency symbols, and Arabic, Hebrew, Hindi and Thai text keeps the marks its script requires; re-import content added through the bulk website, URL or PDF import tools to repair entries indexed earlier.
• Improved: Leftover HTML character codes are converted to readable text before indexing; re-import affected entries to clean up existing ones.
• Fixed: Draft posts published by import tools or scripts are no longer indexed under your homepage address.
• Fixed: WooCommerce prices are indexed in your store’s base currency with the currency code, and on multi-currency stores the chatbot no longer mixes two currencies in one reply; re-import products to update older entries.
• New: The Notification Email field accepts up to five comma-separated addresses, and an invalid address is refused with a message naming it instead of silently breaking notifications.
• Improved: Visitor names and emails from the pre-chat form move into MxChat’s sessions storage automatically — the Leads tab and live-agent greetings work exactly as before.
• Improved: Link-click records are covered by WordPress’s privacy export and erasure tools, and old records anonymize and delete themselves over time.
• Fixed: A YouTube video card now appears only when the video actually matches the question, with a new match threshold setting — and video cards can be switched off entirely.
• New: Transcripts show which AI Tools ran on each message, with timings and errors; tools handling money, customer records or handoffs log only that they ran.
• Improved: The Slack and Telegram handoff tools arrive with their usage note prefilled, so automatic handoff works without extra setup; you can reword or clear it.
• Fixed: Inspecting a Pinecone knowledge entry works, and deletes on Pinecone setups using a namespace remove the right copy.
• Improved: Conversations are stored once instead of duplicated into the options table, and leftover duplicates clean themselves up in the background.
• Fixed: Shared admin features work again on the Transcripts and License pages, and two pieces of admin text no longer point at settings that do not exist.

3.2.18 – August 8, 2026

• Fixed: OpenAI Vector Store knowledge base search works again — an outdated request parameter caused every search to fail silently, so the chatbot answered without your knowledge base content.
• Fixed: The knowledge base status readout now correctly shows how many vector stores are configured instead of always reporting none.
• Fixed: Content updates made outside the editor are no longer skipped. When a post was published by an import tool, a scheduled post going live, or a script, the first edit after it was silently left out of the knowledge base, so the chatbot kept answering from the original version.
• Fixed: PDFs shared in chat no longer store unreadable pages when an embedding request fails part-way through. Failed pages are skipped and recorded in the Debug Mode log, and if no page can be read the chatbot now gives the real reason instead of calling the file empty or unsupported.
• Fixed: Clearing a conversation, deleting a transcript, or erasing a visitor’s data through WordPress’s privacy tools now removes every trace of that session’s stored state. Some of it was previously left behind forever.
• Fixed: Job listings imported from the content selector now include their location, job type and company name, matching listings saved in the editor. Listings imported earlier keep their old text until you import them again.
• Fixed: Post titles are now added to the knowledge base exactly as written. Curly quotes, apostrophes and long dashes were stored as raw character codes, and protected or private posts gained a Protected prefix — all of which could surface in answers. Entries indexed earlier keep their old titles until you import them again.
• Fixed: A short description made up only of page-builder shortcodes or spaces no longer adds an empty Short Description heading to the indexed text.
• Improved: Custom field names containing hyphens now read as words in the indexed text on both import routes, instead of only one of them.
• Improved: Chat sessions no longer bloat your site’s options table. Per-session details were saved as settings rows that were never cleaned up, so busy sites accumulated thousands, some loading on every page view. Sessions now live in their own table, and existing rows are moved across automatically in the background after you update.
• New: Old chat sessions are tidied up on their own. Sessions untouched for 30 days are removed automatically — the window is adjustable with a developer filter — and the cleanup runs even on sites with WordPress scheduled tasks disabled.
• Fixed: Sitemap imports now tell you why they failed instead of pointing at an empty status area — for example, an API key missing the required permission is now named as the reason.
• Improved: Knowledge base import errors from OpenAI, Gemini and Voyage now pass along the provider’s own explanation instead of a generic invalid key message.
• Improved: Chat-side embedding failures are now recorded in the Debug Mode log with the provider’s real reason, matching knowledge base imports.
• Fixed: In the content selector, a selected item’s tick mark was white on a white box and effectively invisible. Selected items now show a clearly filled checkbox.

3.2.17 – August 5, 2026

• Fixed: Japanese, Chinese and Thai content is no longer erased, split into oversized chunks, or stripped of rare characters during knowledge base imports.
• Fixed: Custom embedding providers such as Ollama, LM Studio and Azure now work end to end — the permanent mismatch banner is gone, and no cloud API key is needed for imports or automatic post syncing.
• Fixed: Scheduled posts now enter the knowledge base the moment they go live, and the developer hook for customizing indexed content also runs during automatic post syncing.
• Fixed: The WordPress content import window now fits phone screens, so its buttons are always reachable.
• Fixed: The hybrid keyword boost and smart asset loading toggles now show the saved confirmation like every other setting.
• New: Every settings tab has its own shareable link that opens directly on that tab, including nested and add-on tabs.
• Improved: Knowledge settings are centralized under Knowledge — ACF PDF extraction is now a persistent setting on the ACF Fields tab, and the hybrid keyword boost lives on the new Chunking and Retrieval tab. Saved choices carry over.
• Improved: The Editor Assistant toggle moved to the Content page’s Settings tab, next to the other content tools.
• Improved: The FAQ now documents custom and local embedding providers such as Ollama and LM Studio.
• Improved: Settings holding a Pinecone API key are no longer loaded into memory on every page view.