Gaurav TiwariChangelogPlugin changelog
A site-content chatbot with cited sources, configurable providers, and a compact chat widget.
Fixed: right after updating to 0.6.0, the upgrade could run on many requests at once and queue the background relearn many times. The upgrade and the relearn now run once, under a lock, and a new relearn replaces one still waiting.
New: How Chat answers. Answer length, topic scope (including "only what this site's content covers"), clarifying questions, pointing to the contact form, and up to 20 rules of your own.
ALPHA_CHAT_CONFIG in wp-config.php.Response length budget doubled to 1600 tokens. The 800-token default could cut answers off mid-sentence, which read as a provider failure. New installs now default to 1600; existing installs keep their configured value under Settings → AI…
Launcher offset. Four new fields under Settings → Launcher set how far the floating button and panel sit from the bottom and side of the screen, in pixels, with separate values for desktop and mobile. The launcher was pinned 20px from the…
Fatal error when the plugin update cache was cleared. clear_update_cache() is hooked on delete_site_transient_update_plugins and deleted that same transient, which fired the action again; the hook re-entered itself until PHP exhausted…
clear_update_cache() is hooked on delete_site_transient_update_plugins and deleted that same transient, which fired the action again; the hook re-entered itself until PHP exhausted memory ("Allowed memory size exhausted"). Any code path that deletes the update_plugins site transient triggered it, including wp plugin list after a cache flush. The method now holds a re-entry guard.Error log. Errors and warnings are recorded to a new table and shown under Alpha Chat → Logs, with level filter, expandable context, source attribution, and a clear button. Provider failures used to be invisible unless the site owner read…
alpha_chat_pre_llm_provider and alpha_chat_pre_embedding_provider filters, which replace a provider before the built-in one is constructed. The existing alpha_chat_llm_provider filter can only decorate an instance that already exists, so swapping in your own provider previously still required configuring an API key for a built-in provider you were not using.alpha_chat_log_retention_days, alpha_chat_log_max_rows and alpha_chat_thread_titled hooks.0 messages and a bare uuid in the admin whenever a reply failed. The message count and title were only written after a successful completion, even though the visitor's message had already been stored. Both are now written when the message is stored, and the count is derived from the stored rows.… entity instead of an ellipsis.alpha_chat_logs table and a title_generated column on alpha_chat_threads. Applied automatically on upgrade.Assistant replies showed raw Markdown. Models answer in Markdown, but the widget rendered the reply as plain text, so visitors saw literal bold and - bullets instead of formatting. Replies now render bold, italic, inline code, fenced…
**bold** and - bullets instead of formatting. Replies now render bold, italic, inline code, fenced code blocks, links, headings, and bullet and numbered lists.dangerouslySetInnerHTML. Links are restricted to http, https and mailto; a javascript: URL renders as plain text and every link carries rel="noopener noreferrer".Chat failed with every reasoning model. The stored temperature was sent on each request, and reasoning models (gpt-5*, o1*, o3*, o4*) reject any sampling value other than the default, failing the whole request with 'temperature' does not…
gpt-5*, o1*, o3*, o4*) reject any sampling value other than the default, failing the whole request with 'temperature' does not support 0.7 with this model. temperature, top_p, presence_penalty and frequency_penalty are now omitted for those models instead of being sent with a value they refuse. Conventional models such as gpt-4o are unaffected and still receive the configured values.alpha_chat_is_reasoning_model filter, for a model the bundled prefix list does not cover yet.Security and correctness release from a full audit of the plugin. Please read the upgrade notes below if your site sits behind a proxy other than Cloudflare, or if anything calls the contact endpoint directly.
X-Forwarded-For from arbitrary callers. The header was accepted unconditionally, so rotating it gave every request its own bucket and left provider spend effectively uncapped. Forwarding headers are now honoured only when the request actually arrives from a trusted proxy. Cloudflare's published edge ranges are trusted out of the box; anything else is declared with the ALPHA_CHAT_TRUSTED_PROXIES constant or the alpha_chat_trusted_proxies filter.POST /contact now requires a valid wp_rest nonce. It was previously open to unauthenticated callers and sent mail on every request.origin_url with an attacker-chosen origin_title could previously write arbitrary text into the system prompt.delete_post no longer runs knowledge-base cleanup for revisions and autosaves.WP_DEBUG.wp_safe_remote_request, and the pre-flight check covers IPv6 and literal addresses.alpha_chat_trusted_proxies, alpha_chat_trust_cloudflare, alpha_chat_rate_limit, alpha_chat_max_prompt_faqs, alpha_chat_fallback_candidates, alpha_chat_max_queued_posts and alpha_chat_should_log filters.GET /nonce for widgets on cached pages.define( 'ALPHA_CHAT_TRUSTED_PROXIES', '10.0.0.0/8' );/alpha-chat/v1/contact must now send an X-WP-Nonce header.ChatController::is_rate_limited() is deprecated in favour of AlphaChat\Support\RateLimiter::hit().Public chat no longer dies with Unexpected token '<' when /wp-json/ returns HTML. The widget retries ?rest_route=/alpha-chat/v1/chat and shows a retry message instead of the fallback answer plus a parse dump.
Unexpected token '<' when /wp-json/ returns HTML. The widget retries ?rest_route=/alpha-chat/v1/chat and shows a retry message instead of the fallback answer plus a parse dump.GET /dashboard request. Queue polling only runs while jobs are active and the tab is visible.Curated Q&A can be imported from page URLs. The plugin prefers the local post, then the WordPress REST API (/wp-json/wp/v2/pages and /posts), then HTML. It reads FAQPage JSON-LD, Rank Math / Yoast FAQ blocks, <details>, definition…
/wp-json/wp/v2/pages and /posts), then HTML. It reads FAQPage JSON-LD, Rank Math / Yoast FAQ blocks, <details>, definition lists, accordion markup, and headings that end with ?.POST /faqs/preview and POST /faqs/import (manage_options).alpha_chat_faq_extracted and alpha_chat_page_context filters.wp_rest nonce so WordPress cookie authentication no longer returns "Cookie check failed".www, the static front page, and path fallbacks. Retrieval prefers chunks from that post.Streaming chat at POST /alpha-chat/v1/chat/stream (SSE). The widget streams tokens and falls back to JSON /chat if the host buffers the stream.
POST /alpha-chat/v1/chat/stream (SSE). The widget streams tokens and falls back to JSON /chat if the host buffers the stream.retrieved set.off / low / medium / high) for OpenAI, Grok, and DeepSeek. Default and all presets are low. xhigh/max are not sent.voyage-4-lite, voyage-4, voyage-4-large). Chat, embeddings, and moderation are now separate settings.alpha_chat_moderation_provider filter and a no-op moderator when no OpenAI key is set.ModelCatalog as the admin source of truth. GET /settings now returns { settings, stats, catalog }.xai_api_key and deepseek_api_key.alpha_chat_model_catalog and alpha_chat_http_timeout filters.gpt-5.6-luna.gpt-5.4-mini, gpt-5.4, gpt-4.1) remap on settings read so the dropdown is never blank; the new ID is persisted on Save./chat and /chat/stream require the alpha_chat_frontend nonce the widget already localizes.Added
Changed
url_to_postid()) into the system prompt so "explain this", "summarize this article", and similar queries work without vector-store hits.Threads now capture the originating page URL (origin_url) when a conversation begins. Frontend widget sends window.location.href with the first chat request; admin Conversations tab displays it as a clickable link in the detail panel; CSV…
origin_url) when a conversation begins. Frontend widget sends window.location.href with the first chat request; admin Conversations tab displays it as a clickable link in the detail panel; CSV export includes it as a new column.origin_url VARCHAR(500) column on wp_alpha_chat_threads. Schema version bumped to 1.4.0 — dbDelta adds the column automatically on existing installs.Removed the hard "no retrieval → fallback" gate in ChatService. The LLM is now always called and instructed to handle conversational messages (greetings, thanks, goodbyes, small talk) naturally, reserving the configured fallback message…
ChatService. The LLM is now always called and instructed to handle conversational messages (greetings, thanks, goodbyes, small talk) naturally, reserving the configured fallback message verbatim only for genuine unanswerable factual questions.Frontend widget bundles React directly instead of relying on WordPress core's wp-element / wp-dom-ready script handles. The widget no longer enqueues react, react-dom, react-jsx-runtime, wp-element, wp-i18n, wp-hooks, wp-dom-ready, or…
wp-element / wp-dom-ready script handles. The widget no longer enqueues react, react-dom, react-jsx-runtime, wp-element, wp-i18n, wp-hooks, wp-dom-ready, or wp-escape-html on the frontend.Q&A admin tab for curated question/answer pairs. New wp_alpha_chat_faqs table, REST CRUD at /faqs, and context-injection into every chat request.
wp_alpha_chat_faqs table, REST CRUD at /faqs, and context-injection into every chat request.utm_source=alpha_chat, utm_medium=chat_widget, utm_campaign=ai_answer, utm_referrer={host}).Indexer::index_post now stores a SHA-256 content hash + embedding model per post and skips re-embedding when nothing has changed. Makes "Reindex all" idempotent and cheap.
Indexer::index_post now stores a SHA-256 content hash + embedding model per post and skips re-embedding when nothing has changed. Makes "Reindex all" idempotent and cheap.GET /knowledge-base/queue endpoint returning pending / in_progress / complete / failed counts for the alpha-chat Action Scheduler group.POST /knowledge-base/queue endpoint runs one Action Scheduler batch synchronously (useful when DISABLE_WP_CRON is set).all: initial root. Added a global svg rule + sized close icon; close is now a circular button in the header top-right.:host reset for all SVGs globally, ensuring any icon placed in the widget renders at its CSS size.Initial public release.
[alpha_chat] shortcode; floating launcher is opt-in.alpha_chat_indexable_post_types).