Skip to main content

Changelog

Recent updates and improvements to WizChat.

August 2026​

2026-08-13​

New Features

  • Native (in-page) embed rendering — The script embed can now render the chat panel directly in your page inside a Shadow DOM instead of an iframe, with renderMode: "native". The native panel streams answers from the chatbot's public chat endpoint — no iframe and no secret key in the browser — and composes with both the corner bubble and the docked sidebar. The default iframe embed is unchanged.
  • Connect OAuth MCP servers from the chat — Visitors can now connect and disconnect OAuth-based MCP servers — for example a Microsoft / SharePoint server — directly from the chat's MCP menu, signing in through a popup with a full-page redirect fallback. Each visitor authorizes with their own account, and a server can be limited to specific users or groups.

Bug Fixes

  • Group-restricted MCP servers now appear and connect — An MCP server limited to an allowed user group could be filtered out for every visitor, so it never showed up or connected. Group membership now resolves correctly, and the "Connect" sign-in no longer fails on a redirecting host.
  • MCP tool answers are no longer dropped for dynamically-discovered servers — For MCP servers whose tool catalog is discovered at connect time, a tool's result could be left unattributed and ignored, causing the chatbot to answer as if the tool had returned nothing. Results from these servers are now attributed to the live connection, so the answer reflects what the tool actually returned.

2026-08-12​

New Features

  • Ask about an audio file in chat — Visitors can attach an audio file (MP3 or WAV, up to 25 MB) from the composer's + menu and ask about it. When the chatbot's BYOI chat endpoint is audio-capable, the model listens to the clip directly; otherwise WizChat transcribes it and answers from the transcript. A transcript is carried into the conversation, so follow-up questions run through the normal knowledge-base pipeline with the audio as context. On zero-data-retention / EU-residency chatbots, audio requires self-hosted speech-to-text and text endpoints.
  • Ask about a video in chat — Visitors can attach a video (MP4, WebM, or MOV, up to 100 MB) and ask about it. Video understanding is BYOI-only: it works when the chatbot's BYOI chat endpoint is a self-hosted, video-capable model, which watches the video through a short-lived signed URL. Chatbots without a video-capable endpoint return a clear message instead.

2026-08-10​

New Features

  • Docked sidebar embed mode — The script embed can now render the chatbot as a full-height right rail that pushes your page content aside when it opens, instead of the corner bubble. Opt in with displayMode: "sidebar"; you can point it at the element to reflow, hide the corner launcher, and open it from your own button via a data-wizchat-open attribute or Ctrl/Cmd + /. The default bubble embed is unchanged.
  • Restrict an MCP server to specific embedding sites — When the same chatbot is embedded on more than one website, an individual MCP server can now be pinned to a list of allowed embedding domains, so its tools are exposed only on the sites you trust and hidden everywhere else. Servers with no list stay available everywhere (unchanged); once a list is set, requests from unknown hosts are fail-closed.

2026-08-09​

New Features

  • Pick your speech language for voice input — The in-chat microphone's options popover now has a Speech language picker. The Web Speech API recognizes one language per session and cannot auto-detect, so choosing a language means speech in that language (for example Hebrew) is transcribed correctly instead of coming out as English gibberish. The choice is remembered per browser; it also passes a language hint to server transcription, but only when explicitly chosen, so the auto-detecting fallback is unchanged for visitors who never open the picker.

Improvements

  • Owner config changes apply to open chats automatically — An already-open chat tab now picks up an owner's configuration changes the next time the visitor returns to that tab, without a manual page reload. The refresh is silent — it never flickers the microphone's loading state or replaces good config on a transient network error.

Bug Fixes

  • Slow self-hosted models no longer time out mid-answer — A BYOI query against a slow or cold-starting self-hosted model could fail after about two minutes — even while the model was still working — because the chat request hit its time limit. WizChat now allows a chat to run up to five minutes end-to-end and gives each call to your endpoint up to 90 seconds, so a slow or cold-starting model has time to respond. On very slow endpoints, WizChat also bounds how many reasoning steps a single answer takes, so the reply still comes back within the limit.

2026-08-08​

New Features

  • Transcribe voice input through your own endpoint — For chatbots using BYOI, when your endpoint configuration routes speech-to-text to one of your own endpoints, the in-chat microphone now transcribes through your model instead of the platform's default speech-to-text service. Audio is relayed to your endpoint server-side — it still transits WizChat but is only ever sent to your own model, never to the platform's speech-to-text provider. Transcription is fail-closed — it never silently falls back to the platform service.

Improvements

  • Owner-configurable AI-action limits for MCP dialogs — In-dialog wizchat/llm completions and model-backed tools are now bounded by a per-user, per-chatbot rate limit across three fixed windows (per minute / hour / day; defaults 10 / 60 / 300, 0 = unlimited) that owners can tune per chatbot, replacing the previous fixed 5-completions-per-dialog cap. Hitting a limit returns a clear 429 that names the limit and how long to wait.

Bug Fixes

  • MCP dialogs no longer dead-end at the action limit — Reaching an MCP-UI dialog's action limit now returns a clear, recoverable error instead of breaking the dialog with an "Unknown UI resource" failure that forced a reopen.

July 2026​

2026-07-30​

Improvements

  • Sharper source selection and citations — The sources shown with an answer are now ranked and filtered by a keyword-free relevance ("meaning") score, so the most on-topic sources appear first and weak keyword-only matches no longer take up slots. Citations into multi-page PDFs open the page that best fits your question, preferring the start of the relevant section when two pages are a near tie.
  • More relevant, less redundant sources — Retrieval now blends keyword and semantic matching based on how a question reads and merges multiple chunks from the same document into a single source, so distinct sources surface instead of being crowded out by repeats of the same file.
  • Balanced Sources surfaces on-topic videos in the answer — When Balanced Sources is enabled and a genuinely on-topic video is retrieved, the chatbot now reads that video and factors it into the answer, so a relevant video reaches both the answer and the displayed sources instead of being crowded out by higher-scoring text. No video is forced when none is on-topic.
  • Video sources deep-link to the right moment — Video source cards now show the timestamp of the cited moment and open the video at that point when clicked, including for older transcripts that store times as [MM:SS].
  • Clearer text-ingestion error responses — When every entry in a POST /api/v1/chatbots/{chatbotId}/text request fails, the 500 response now returns a structured error object (code: "ingest_failed" with a human-readable message) alongside the per-entry documents[] reasons, so failures are easier to diagnose programmatically.

Bug Fixes

  • No more theme flash on reload — The chat page no longer briefly flashes the light theme before applying a saved dark preference when reloading.
  • No more broken images in answers — When an answer references an image that does not actually exist in the knowledge base, the chatbot now removes the broken image instead of showing a broken-image icon, keeping only the image's caption text. Images that do exist are unaffected, and a temporary storage hiccup never causes a valid image to be dropped.

2026-07-28​

Improvements

  • Ingest YouTube videos by link — POST /api/v1/chatbots/{chatbotId}/youtube/videos now accepts a links array (watch, youtu.be, embed, or shorts URLs, or bare 11-character IDs) alongside videoIds, so you can ingest a video even when it is not shown in the channel list. Provide videoIds and/or links; any link that cannot be parsed to a video ID is ignored and returned in the response's invalidLinks array.

2026-07-27​

New Features

  • Restrict web search to specific domains — When web search is enabled, you can now provide an allowlist of up to 20 domains so the chatbot only pulls live results from sources you trust. Leaving it empty keeps web search unrestricted (the default).
  • Hide the default scope from the selector — Chatbots with knowledge scopes can now hide the default (base-namespace) entry from the scope picker so end users only see your named scopes. Exposed on the Management API via showDefaultScopeInSelector on GET/PATCH /api/v1/chatbots/{chatbotId}/scopes.

2026-07-25​

New Features

  • Add knowledge base content via the API — A new POST /api/v1/chatbots/{chatbotId}/text endpoint adds text or Markdown straight to a chatbot's knowledge base with no file upload. Send a single text (with an optional fileName) or a batch of up to 50 items — for example curated question/answer pairs. Each entry becomes its own searchable, listable, deletable source; processing is synchronous and returns per-entry results (207 on partial success). Supports viewAccessLevel, authorizedEmails/authorizedGroups, and scopeId. Requires the documents:write scope.

2026-07-19​

New Features

  • Analytics API — Rich per-category analytics endpoints under GET /api/v1/chatbots/{chatbotId}/analytics/*: queries, quality, conversations, search, tokens, infra, and transcription (alongside the existing usage and cost). A new workspace-level GET /api/v1/analytics/aggregate returns a usage overview across all chatbots. All require the analytics:read scope.
  • Scheduled Crawls API — Full CRUD for scheduled website crawls: GET/POST /api/v1/chatbots/{chatbotId}/scheduled-crawls and GET/PATCH/DELETE .../{scheduleId}. Knowledge-base infrastructure (region and index) is derived server-side. Requires documents:read / documents:write.

Improvements

  • Quality analytics answerRate — Now returns null instead of 0 when a period has no classified conversations, so "no data yet" is distinguishable from a genuine zero. Label-analytics count denominators were also corrected.

2026-07-16​

New Features

  • Knowledge Scopes API — Per-scope endpoints POST /api/v1/chatbots/{chatbotId}/scopes, GET/PATCH/DELETE .../scopes/{scopeId}, complementing the existing list and bulk-update calls. Requires chatbots:read / chatbots:write.

Improvements

  • More reliable math — Numeric questions now run through a dedicated run_python compute tool with LaTeX-rendered formulas, and pure-math queries are routed straight to it. This fixes cases where calculations were approximated or formatting was inconsistent.
  • Steadier tool use — When a connected MCP tool is called repeatedly with identical arguments, the chatbot now breaks the loop and routes to Code Mode instead of stalling.
  • Faster, clearer response preparation — Pre-agent setup steps now run concurrently and show descriptive progress (e.g. "Searching knowledge base"), so users can see what's happening during longer turns.

2026-07-14​

Bug Fixes

  • Document uploads now self-heal by creating the storage bucket on demand if it is missing, preventing upload failures for newly deployed chatbots.

2026-07-13​

Improvements

  • Faster responses on strong-RAG turns — The chatbot now skips the quality-evaluation step when the retrieval engine returns a strong, unambiguous match. This reduces response latency by up to 5–6 seconds on those turns with no change to answer quality. The improvement applies automatically to all chatbots with no configuration required; chatbot owners can override the behavior per deployment via the EVALUATOR_SKIP_MODE environment variable.

2026-07-12​

New Features

  • Sales Agent — Single Lead Detail (PII) — New GET /api/v1/chatbots/{chatbotId}/sales-agent/leads/{leadId} endpoint returns full decrypted contact PII (name, email, phone, company, custom fields) plus coarse geo for one lead. Audit-logged. Requires security:read scope (owner/admin only).
  • Sales Agent — Leads CSV Export — New GET /api/v1/chatbots/{chatbotId}/sales-agent/leads/export endpoint exports up to 5,000 leads as a CSV with decrypted contact fields and a pii_decrypt_status column per row. Requires security:read scope (owner/admin only).
  • Sales Agent — GDPR DSAR Lead Erasure — New POST /api/v1/chatbots/{chatbotId}/sales-agent/leads/{leadId}/dsar-erase endpoint hard-erases a lead's PII while retaining the anonymous row for aggregate metrics. Idempotent; repeat erasure returns 409 with original piiErasedAt preserved. Requires security:write scope (owner/admin only).
  • Chatbot Users API — Three new endpoints manage deployed end-users: GET /api/v1/chatbots/{chatbotId}/users lists all invited users; POST invites a new user by email (with optional display name and role); DELETE ?userId= revokes access. Requires security:read / security:write scope.
  • Chatbot Settings API — New PATCH /api/v1/chatbots/{chatbotId}/settings endpoint updates runtime toggles (web search, unanswered-question tracking, long-term memory, memory strategy). Destructive changes (disabling memory or switching strategy) require confirmDelete: true. Requires chatbots:write scope.
  • Memory Data Deletion API — New DELETE /api/v1/chatbots/{chatbotId}/memory-data?scope=<scope> permanently clears graph memory by scope (long-term or short-term). Requires chatbots:write scope.
  • Email Groups API — Full CRUD for owner-scoped email allow-list groups: GET /api/v1/email-groups (list, with optional includeUsage), POST (create), GET /{groupId} (fetch), PUT /{groupId} (full replace), DELETE /{groupId}. Requires security:read / security:write scope.

2026-07-11​

Improvements

  • ZDR final-answer routing — Final answers are now routed through Zero Data Retention-compliant providers when ZDR mode is active, ensuring the complete response pipeline is ZDR-compliant end-to-end.
  • Per-agent evaluator reasoning effort — Each agent's evaluator call now uses the reasoning effort configured for that agent rather than a fixed default, giving chatbot owners finer control over quality vs. cost trade-offs per scope.

Bug Fixes

  • Scope submenu now expands inline on narrow viewports instead of overlapping the parent menu (#445).
  • Scope menu no longer opens centered and jumps left on first render (#444).
  • return_script MCP servers now work correctly alongside remote_http MCP servers on the same chatbot — mixed-mode configurations no longer error (#446).

2026-07-09​

New Features

  • Sales Agent API — New endpoints manage the Sales Agent via the Management API: GET /PATCH /api/v1/chatbots/{chatbotId}/sales-agent reads and updates the configuration (prompt overlay, model override, lead-capture fields, guardrail, trigger sensitivity); GET /api/v1/chatbots/{chatbotId}/sales-agent/handoffs lists lead-delivery channels; PATCH /DELETE /api/v1/chatbots/{chatbotId}/sales-agent/handoffs/{handoffId} updates or removes a channel. Requires chatbots:read / chatbots:write scope.
  • YouTube Management API — Four new endpoints automate YouTube ingestion: GET /api/v1/youtube/status (connection status), POST /api/v1/youtube/connect (start OAuth flow), GET /api/v1/youtube/videos (list channel videos with pagination and title search), and POST /api/v1/chatbots/{chatbotId}/youtube/videos (ingest videos into the knowledge base). Requires videos:read / videos:write scope.
  • Transcript Corrections API — Three new endpoints expose the AI-generated transcript correction workflow: GET /api/v1/chatbots/{chatbotId}/corrections returns a chatbot-wide review queue; GET /api/v1/videos/{videoId}/corrections returns paginated corrections for one video (with context snippets); PATCH /api/v1/videos/{videoId}/corrections resolves a correction (accept candidate, reject, apply custom text, or undo — accepted or custom corrections trigger automatic re-embedding). Requires documents:read / documents:write scope.

Improvements

  • Persistent pre-indexed MCP tool vectors — MCP tool embeddings are now computed and stored at index time, significantly reducing latency for tool discovery on large MCP servers.

2026-07-08​

New Features

  • Delete a chatbot API — New DELETE /api/v1/chatbots/{chatbotId} endpoint permanently deletes a chatbot and cascades cleanup across the knowledge base, storage, deployment, and configuration. Requires the chatbot owner or a team owner/admin (chatbots:write scope). Returns 200 on full deletion, 207 Multi-Status if the chatbot was removed but some downstream cleanup failed, or 409 if a shared resource (e.g. a knowledge base used by another chatbot) blocks deletion. The error schema gains two new codes: conflict and deletion_failed.
  • Document Ingestion API — New POST /api/v1/chatbots/{chatbotId}/documents endpoint ingests a PDF into the chatbot's knowledge base directly from a public URL. The PDF is fetched, extracted, chunked, and embedded asynchronously; poll GET /documents for status. Requires documents:write scope. SSRF-safe: http(s)-only, public-IP-pinned, size- and time-bounded.

2026-07-07​

New Features

  • Website Crawl Management API — Three new endpoints bring the dashboard website-crawl flow to the Management API:
    • POST /api/v1/chatbots/{chatbotId}/crawls — start a crawl for one URL or a batch of up to 100 URLs; the chatbot's knowledge base is updated with the crawled content as vectors.
    • GET /api/v1/chatbots/{chatbotId}/crawls — list crawls (most recent first) with live progress; non-terminal crawls are refreshed against the crawler on read.
    • GET /api/v1/chatbots/{chatbotId}/crawls/drift — compare the chatbot's crawled URL set against the domain's current sitemap.xml to surface pages added or removed since the last crawl.
  • Team Members API — New GET /api/v1/team endpoint lists the account owner's team members.

2026-07-06​

New Features

  • Sales-agent leads API — GET /api/v1/chatbots/{chatbotId}/sales-agent/leads lists captured lead metadata (disposition, handoff status, country code, timestamps, and aggregates) without exposing PII. Requires analytics:read scope; owner/admin only.
  • User session credential source — Embedded chatbots can adopt an authenticated host-dashboard session via a custom token, enabling seamless single-sign-on for co-pilot integrations.
  • New Chat button in embed widget — Users can start a fresh conversation from within the embedded widget without refreshing the page.

2026-07-05​

Improvements

  • Per-MCP-server reasoning effort — Each MCP server's model override now accepts a reasoningEffort field (none / minimal / low / medium / high / default) to set thinking intensity independently per server.
  • Wide markdown tables in the chat widget now scroll horizontally instead of collapsing.
  • <br> tags now render correctly inside markdown table cells.
  • @provider model pins fall back gracefully when the requested provider variant is unavailable.
  • MCP servers that return tool calls as plain text are handled without errors.

2026-07-01​

New Features

  • Tool approval policies — Chatbot owners can configure per-tool approval policies for MCP servers. Each tool can require user confirmation before the AI acts. Users can grant persistent "Allow always" approval from within the dialog without leaving the chat.

June 2026​

2026-06-30​

Improvements

  • MCP tool results are now replayed to the model across conversation turns, ensuring consistent tool context in multi-turn sessions.
  • Code-mode agent no longer enters a loop when a change is already staged for the current step.
  • Graceful cancellation when a user denies a tool call mid-flow in code mode.

March 2026​

2026-03-03​

New Features

  • In-chat data analysis — End users can upload Excel and CSV files directly in the chat for instant analysis. Ask questions in natural language and get answers with interactive charts and tables.
  • Knowledge Scopes — Partition your knowledge base into named scopes with independent access control, behavior overrides, MCP server filtering, and budget caps (Business & Enterprise).
  • Per-plan data retention — Chat history, Q&A logs, and in-chat file uploads are now retained based on your subscription plan, from 30 days (Free) to unlimited (Enterprise).
  • Per-scope cost limits — Set monthly or weekly budget caps on individual knowledge scopes to control spending.

Improvements

  • Smarter sheet selection for multi-sheet Excel workbooks — the AI automatically picks the most relevant sheet based on your question.
  • Follow-up questions remember context (selected sheet, metrics) across the conversation, so you don't need to repeat details.
  • Automatic chart type selection based on question intent — trend questions get line charts, comparisons get bar charts.
  • Web search is automatically disabled when querying internal data (spreadsheets, databases) to avoid irrelevant external results.

February 2026​

2026-02-25​

New Features

  • Image understanding in chat — Upload images directly in the chat and ask questions about them. The AI analyzes your image and answers your question directly.
  • Network visibility on user rows — For open-access chatbots, each user row now shows a masked network badge, question count, and last-seen timestamp.

Improvements

  • Faster responses to image questions — roughly 2x faster than before.
  • Analytics data is now always up to date when you open the page.

2026-02-23​

New Features

  • Per-metric access control — restrict individual dashboard metrics to specific users or email groups
  • Web search toggle — enable or disable Perplexity web search per chatbot from the Settings tab
  • Domain-based access control — restrict chatbot access to specific email domains
  • PostgreSQL metrics dashboard — auto-detected metrics and custom metric charts for your databases
  • Multiple PostgreSQL schema support — select multiple schemas per connection

Improvements

  • MCP servers can now be edited after creation (update credentials, endpoints, or settings in place)
  • PostgreSQL metrics panel supports multiple database connections with clear server labels
  • CHM upload hard cap increased from 250 MB to 350 MB
  • Video uploads now use signed URLs for improved security

Bug Fixes

  • Fixed mixed-case PostgreSQL table names not being detected in the metric wizard
  • Fixed duplicate server names when multiple PostgreSQL databases share a host
  • Fixed video upload progress bar jumping backward during processing

February 2024​

2024-02-01​

New Features

  • Added Zero Data Retention (ZDR) mode for maximum privacy
  • Cerebras model routing for fast ZDR-compatible inference
  • Cost tracking across all AI usage

Improvements

  • Improved document processing speed by 40%
  • Better error messages for failed uploads
  • Enhanced search relevance

Bug Fixes

  • Fixed OAuth token refresh for Atlassian integration
  • Fixed team invitation emails not sending
  • Fixed analytics date range picker

January 2024​

2024-01-15​

New Features

  • Team billing with ownership transfer
  • Share request workflow for non-owners
  • Scheduled website crawling

Improvements

  • Redesigned chatbot settings UI
  • Faster deployment times
  • Improved mobile responsiveness

Bug Fixes

  • Fixed PDF processing for certain encrypted files
  • Fixed conversation history not loading
  • Fixed duplicate document upload detection

2024-01-01​

New Features

  • MCP server support (Salesforce, Atlassian, PostgreSQL)
  • Custom MCP server integration
  • Document access control by email

Improvements

  • New analytics dashboard
  • Improved embedding quality
  • Better handling of large documents

December 2023​

2023-12-15​

New Features

  • YouTube channel integration
  • Video transcription improvements
  • Bulk document upload

Improvements

  • Faster knowledge base search
  • Improved chat UI
  • Better mobile experience

Bug Fixes

  • Fixed team member permissions
  • Fixed export functionality
  • Fixed timezone issues in analytics

2023-12-01​

New Features

  • Team collaboration features
  • Role-based access control
  • Chatbot sharing

Improvements

  • Redesigned dashboard
  • Improved onboarding flow
  • Better documentation

Earlier Releases​

For release notes prior to December 2023, please contact support.

Upcoming​

We're working on:

  • Additional AI model options
  • Enhanced analytics
  • More MCP integrations
  • Improved embedding models

Stay tuned for updates!

Feedback​

Have suggestions for new features? We'd love to hear from you: