Changelog

​

Knowledge

• An experiment scores each retriever on Recall, Precision, MRR, and Hit Rate. You set the weight of each metric.
• An experiment card shows the three best retrievers of the run.
• The question detail compares the expected answer with the retrieved text and marks the missing parts.
• When your default retriever already scores best, the result says that there is nothing to change.
• Run experiment stays disabled until the knowledge base has a retriever that is ready to answer.
• Fixed: a fully retrieved answer is marked found and gets correct metric scores. Run the experiment again to update an older result.
• Fixed: the best retriever of a run always shows in the ranking.

​

Agent builder

• Fixed: the help panel of the Custom Code gambit reads variables through variables, such as variables.response.Name. See Custom Code.

​

Data residency

• India joins the United States, the European Union, and Qatar as a data region. See Organization lifecycle.

​

Campaigns

• A campaign reply starts a new conversation even when one is already running, including while a human agent is in control. See Handle replies.
• The reply that started the conversation is kept in trigger.userMessage.

​

WhatsApp

• List and rating buttons read Show options and Choose a rating, and rating labels show on WhatsApp. See End user experience.
• A message longer than 4,096 characters is split into more than one message.
• Show knowledge citations now applies on WhatsApp.
• A message that could not be delivered shows in the Activity tab. See Watch conversations.

​

Security

• Form answers follow your personal data protection settings. See Personal data protection.

​

Agent builder

• API Call gambits send their headers with the request. See API Call.

​

Knowledge

• A new custom retriever gets a readable name from its search type and embedding model, such as Hybrid · OpenAI · 3-large. You can change the name. See Configure a retriever.

​

Tools and CodeKit

• Array and object inputs in Workflow Mode keep their structure when a referenced value contains quotes or line breaks.
• An input that is not valid JSON, or has the wrong type, stops the action before it runs.

​

Knowledge

• The Dataset tab is redesigned. Each test set shows as a card, and you can set document priority, question types, and dimensions when you generate test questions.
• You can edit a test question and select the exact answer text in its source document.
• Two test sets in one knowledge base cannot share a name.

​

Chat widget

• A video from a knowledge base plays directly in the agent reply.
• Card displays show images from a knowledge base.
• Captions on agent media replies show formatted text.

​

Tools and CodeKit

• CodeKit tool gambits run in Workflow Mode, with explicit inputs and success and failure routes. See Enable and use actions.
• A CodeKit action can be turned on only after it is deployed. The action picker shows the deploy state of each action.
• You can duplicate a CodeKit action or a whole CodeKit toolkit. A new toolkit starts with a Greeting action.
• The schema editor gains Form and JSON views and a Supported schema guide you can copy.
• Change account and Reconnect refresh the existing authorization first. A new authorization tab opens only when the provider rejects the refresh.
• A CodeKit action is retried only when the call failed before its code started, so a retry never runs the code twice.

​

Knowledge

• The Retriever tab shows shared indexes, build progress, and documents under repair.
• Retriever number fields show their allowed range.
• Validation flags a Knowledge Retrieval gambit whose retriever cannot answer yet, before you publish. See Attach knowledge to an agent.

​

Agent builder

• Input gambits gain an Allow Skip switch.
• Condition rules can compare the value of the gambit itself.

​

Chat widget

• A Redirect gambit that opens in a new tab shows a link for the end user to select.
• A resolved conversation leads to a new conversation in the widget.
• Widget Embed has a GTM tab with a Google Tag Manager snippet. See Embed the widget.

​

Live chat

• Active chat counts drop when a response timeout takes a conversation away from a human agent.

​

Data residency

• Qatar joins the United States and the European Union as a data region. See Organization lifecycle.

​

Knowledge

• You can edit a document title and body inline and group documents by source.
• A ready document no longer shows a status badge.

​

Agent builder

• Fixed: notes keep typed text, Ctrl+Shift shortcuts work, and Search Canvas opens again.

​

Security

• Tars support can erase an end user’s personal data on request. See Erasure and legal holds.

​

WhatsApp and campaigns

• Connect WhatsApp number offers two providers, Direct (Meta) and 360dialog. A number you already run on 360dialog connects with its API key. See Connect WhatsApp.
• The number detail sheet shows the provider of each number. For a 360dialog number, it checks where incoming messages go and can point delivery back to Tars.
• An Admin can copy a 360dialog number’s API key from the number detail sheet. Each copy is recorded in the audit log.
• Templates live in Campaigns → Templates and apply to the number you select. See WhatsApp templates.
• A template created from a campaign or the builder uses the number that sends it.
• A number import that fails points to a retry that runs the import again.

​

Security

• Admins can turn on personal data protection in Settings → Privacy. See Personal data protection.
• Protection finds card numbers, Social Security Numbers, email addresses, phone numbers, IBANs, and custom patterns in end user messages, captured fields, and CSAT feedback.
• Redact removes a value permanently. Mask stores a placeholder and keeps an encrypted original for 1 to 365 days.
• Card numbers and Social Security Numbers are always redacted.
• A member in live chat can reveal a masked value for 60 seconds, and each reveal is recorded in the audit log.

​

Agent builder

• Reasoning is off by default unless the model requires it. The Model parameters popover groups its settings and adds Max reasoning tokens.
• A parameter that the selected model does not support stays visible but disabled.

​

Knowledge

• When you upload files, you can pick, filter, and clear them. A file that is too large shows before the upload starts.
• A website import sets max pages and max depth up front. Advanced settings holds Include subdomains and User agent.

​

Billing

• The plan conversation allowance keeps renewing during a payment grace period.

​

Knowledge

• Knowledge bases open in a refreshed list and grid, with descriptions of up to 500 characters, document counts, and a create tile.

​

Agent builder

• The Language gambit gains an Allow Free Text switch.
• Turning Conditions off brings the option handles back.
• Fixed: editing a gambit no longer undoes connections drawn while its panel was open.

​

Fixes

• The conversation data card no longer lists city and country twice.

​

Chat widget

• Auto-read starts off on every visit and applies only to the current agent.

​

Knowledge

• A website import can send a custom User agent.
• Website imports capture content that loads late on JavaScript-rendered sites.

​

Fixes

• The signup phone field preselects your country.

​

Chat widget

• The toolbar on agent replies stays visible and shows the time. Read aloud and copy on the last message cover the whole run of messages.
• Multi-select cards collect picks as chips and send them together.
• Fixed: image galleries show public external image links again.

​

Knowledge

• With Media search on, an agent can find and show images and videos from a knowledge base. Manage media lets you add context to each item.
• Knowledge base videos play in the chat instead of opening as a link.

​

Agent builder

• Two new system variables, system.region and system.timezone, join system.country and system.city in the variable panel. The time zone uses the IANA format.
• Location variables are read when the widget page loads and stored when the conversation is created, so they are set from the first gambit.

​

Agent replies and in-chat input

• With File Reading on, an agent can answer from files the end user sends in the conversation. PDF, Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and plain text files are read. Images are described, with their text transcribed.
• A scanned PDF with no text layer is transcribed, so the agent can still read it. Video and audio files are not read.
• The agent reads a file by section instead of receiving the whole file, so a large document does not crowd out the conversation. Each read shows as a tool call when tool calls are visible.
• Files answered through a Media Upload gambit or a file upload component earlier in the flow are readable by a later AI Agent gambit in the same way.
• A file that cannot be read is reported to the agent with a plain reason, such as a password-protected or damaged file. A file over 10 MB is not read.

​

Chat widget

• Attached files upload as soon as they are added. Each file shows uploading, then its size, or failed with a retry message. Send waits until every upload has finished.
• A message carries at most 10 attachments of up to 10 MB each, and the composer names the file when one is too large.
• Fixed: the thinking indicator shows only while an AI Agent gambit takes its turn, and no longer on other gambits.

​

Agent builder

• A File Reading built-in tool joins the Built-in Tools & Capabilities dialog. It is off by default, like the other built-in tools. See Built-in AI capabilities.
• Fixed: Copy link on an agent card copies a working builder link. The old link opened an Agent not found page.

​

WhatsApp

• A document or image sent to a WhatsApp number is stored by Tars. An agent with File Reading on can answer from it like a widget upload.

​

Live chat

• Fixed: desktop notifications and sounds no longer fire for a conversation that is resolved or that the agent controls again. New assignments still alert you.
• Fixed: the agent response timeout no longer bounces a live chat back to a disconnected human agent every few minutes. It now hands the conversation only to a human agent with the dashboard open.
• When nobody is eligible, the conversation returns to the agent flow or waits unassigned. See Assignment and availability.

​

Security

• The dashboard and the chat host send a Content-Security-Policy header. Dashboard pages cannot be loaded inside an iframe on another site.
• The chat and widget pages still load inside an iframe on any site, so the widget script and the iframe embed keep working. The preview page can be framed only by the chat host itself and the dashboard, and every other chat host page refuses framing.

​

Agent builder

• Fixed: Custom Code gambits run again. Every custom code execution had failed with a sandbox error before the code started.

​

WhatsApp and campaigns

• A number alias holds up to 50 characters, a tag name up to 30, and a number up to 5 tags. Counters in the detail sheet show the current length.
• Length is counted in Unicode code points, so most accented letters and simple emoji count once, while a flag or skin-tone emoji counts more than once.
• Changing or disconnecting a number’s agent asks for confirmation only when an active or paused webhook campaign uses the number. The always-on re-route dialog in Distribute is gone.
• In the agent picker for a number, an agent already used by another number stays listed but disabled, with a Used by note naming that number.
• The number detail sheet is part of the page address. A reload reopens the same sheet, and closing it clears the address so back navigation does not reopen it.
• A template that Tars could not submit to Meta shows Template submission failed with the reason, plus Edit & resubmit and Dismiss failed attempt. A Meta rejection keeps its own Fix & resubmit action.
• A campaign whose header media is unavailable fails with media_missing and asks you to upload the media again before relaunching.
• Campaign delivery milestones are consistent. A delivery that reaches Read or Replied also counts as Delivered, so a reply never appears without a delivery.
• The webhook campaign API Contract card offers an importable Postman Collection and an OpenAPI 3.1.0 document next to the cURL example.
• A webhook campaign call with a recipient outside canonical international format fails with invalid_params and an errors entry of invalid_e164 on the to field.
• A webhook campaign send matches an existing end user by phone number with or without the leading +, instead of creating a duplicate end user.

​

Data residency

• The platform runs regional cells, and an EU cell now joins US. An organization’s data and its end-user traffic stay in that organization’s region.
• Widget and API traffic resolve the region from the embed. An organization that moves region keeps working with the embed code already on your site, with no change needed there.
• A missing or unknown region now fails with a clear error instead of quietly serving another region’s data.
• Knowledge retrieval keeps only vectors and identifiers in the search index. Document content is read from your own region’s database.
• Tars support places a new organization in the EU region and moves an existing one.

​

Agent builder

• The Conditions panel is available on the no-input gambit, with the same branch routing and Default fallback as other gambits.

​

Fixes

• A session that expires while the dashboard is open now returns to sign-in instead of an error screen.
• The error screen gains a Sign out button next to Try again.

​

Agent replies and in-chat input

• Agents can answer with interactive components instead of plain text. The set covers charts, data tables, cards, status lists, metric cards, progress trackers, timelines, accordions, code blocks, image galleries, and maps.
• Agents can collect answers inside the conversation using forms, button groups, confirmation dialogs, calendars, time slots, file uploads, and one-time code entry.
• An answered component freezes in place and keeps showing the answer, including on dashboard activity screens.
• Agent thinking is visible. A thought pill reports how long the agent reasoned, and each tool call shows its status and its payload.
• Agent replies gain a hover toolbar for read aloud, feedback, and copy. A link an end user sends unfurls into a preview card.
• Exit routes are editable. You can add, rename, describe, and remove a route, change where it points, and restore the default Complete route in one action.
• Paid plans can hide the Tars badge and set a footer disclaimer of up to 300 characters. Free organizations always show the badge.
• The escalate-to-human tool is retired, and human handoff now runs through exit routes. The standalone date picker is retired in favor of the calendar component.

​

Chat widget

• The widget opens straight into the chat. The home screen and the bottom tab bar are gone.
• The widget boots in a single round trip and preloads while the browser is idle, so it opens faster.
• A conversations list is reachable from the chat header. Rows are grouped by day, sorted by last activity, and can be swiped away.
• A returning end user can land on the conversations list instead of the chat. Turn on Open to conversations list in Distribute → Widget Embed.
• Spoken replies have a play control under each reply, and playback stops when the conversation changes.
• Fixed: typing indicators leaking between conversations, doubled question bubbles around components, and scroll losing the bottom of a conversation. A reply now quotes the message it answers, so the agent receives the right context.

​

Billing

• A Premium signup now grants the conversation volume it paid for. Some organizations kept the free allotment after upgrading.

​

Knowledge

• A knowledge base can now grade itself. The Dataset tab builds a test set from your documents, and the Experiments tab measures how well each retriever answers it.
• An experiment compares retrievers over one shared test set, so you can weigh recall against precision before you promote a retriever.

​

WhatsApp and campaigns

• Connecting or reconnecting a number shows live progress, keeps a background progress chip, and reports a failure you can retry.
• The Connect and Reconnect actions stay disabled until onboarding is available.
• Template submission is idempotent and shows Pending immediately. Header, body, and footer limits count Unicode code points, so accented text and emoji no longer miscount.
• Media rules are enforced on upload, replacement, submission, activation, and webhook send. Unhealthy and test-only accounts can no longer send a campaign.
• Campaign Activity reconstructs the conversation, pairing one outbound template message with the inbound reply.
• A webhook campaign dry run causes no delivery, no Meta call, and no credit movement.

​

Billing

• Invites are no longer blocked by a stuck seat cap.
• An invoice settled from a credit balance carries its real value instead of showing zero, and Enterprise credit grants complete correctly.

​

Organization and data lifecycle

• Data lifecycle now runs end to end, covering retention by plan, archival, organization-wide legal holds, and authorized archive exports. See Data retention.
• A legal hold blocks permanent deletion and archive expiry, while lossless archival continues.
• Test and preview conversations are permanently deleted after 30 days and are never archived.
• Deleting an organization runs as a bounded, resumable teardown that verifies no data remains before it completes.
• Onboarding no longer fails for a member who belongs to an organization but has not selected one.
• Idle conversations are closed by the abandonment sweep again, which had stopped running.

​

Knowledge

• Indexing no longer fails under load, and a stalled indexing job now recovers on its own.
• Indexing inside one knowledge base runs one document at a time. A single large knowledge base now indexes more slowly, and a faster strategy is planned.

​

Support access transparency

• Tars support access to your organization is now time-limited and visible. A banner shows who has access and the remaining time, and you can end access early.
• Support actions appear in your audit log as Tars Support, with the individual identity kept internally for accountability.

​

Knowledge lifecycle

• Retrievers now show a Needs reindex state with a Reindex all action, so index staleness is visible instead of silent.
• Imported documents can be edited in place, with revision history that records whether content came from an upload, an edit, or a website sync.
• Website re-fetch schedules warn when they cannot apply.

​

Fixes

• Fixed label creation on the conversations page.

​

Chat widget

• The widget’s file upload now enforces the configured maximum file size.

​

Conversations and delivery

• Agent replies, media, human-agent replies, and voice notes now deliver consistently on every active channel.
• Publishing reliability improved: live conversations always run the pinned published flow, never a draft in motion.

​

Campaigns and analytics

• Campaign screens were reworked for clarity, and delivery hardening reduces stuck or duplicated sends.
• Analytics gained drilldowns from summary numbers into the supporting conversations.

​

Tools

• Large toolkits load faster in the tool picker, and their action lists render reliably.
• Tool outputs can be projected to only the fields an agent needs, which keeps AI context small and answers focused.

​

Analytics

• Analytics gained metrics from page load through lead capture.

​

Knowledge and answers

• PDF image extraction is now plan-gated, and knowledge citation visibility is a per-agent design setting.

​

Billing

• Several Stripe purchase edge cases were fixed.

​

Analytics

• Daily analytics rollups now use consistent UTC dates.

​

Billing

• Plan pricing became plan-scoped, so organization-specific terms never leak between customers.

​

Onboarding and setup

• Organizations can upload a logo during onboarding.
• Sign-out is available from the onboarding screen, and invites to existing organizations work reliably.

​

Knowledge and answers

• Website import quality improved for hard-to-scrape sites.

​

Analytics

• Analytics was redesigned with a generic reporting system that covers campaigns alongside conversations.

​

Tars 3.0

• Organizations are now the home for every resource: agents, knowledge bases, tools, channels, campaigns, end users, and billing. Five roles (Admin, Member, Viewer, Billing, Auditor) and separate live-chat seats control access.
• The agent builder gains AI Agent gambits, exit routes, drafts and publishing, versioned rollback, and a test preview that is never billed. A rebuilt conversation engine runs every flow the same way across channels.
• Conversation variables move to ten clear namespaces with a pill-based insert menu. A test request can load real sample data, so field autocomplete works from actual responses.
• Tools become a first-class system: a directory of connected toolkits with OAuth and API-key connections, per-action enablement, and tool gambits with success and failure routes on the flow.
• CodeKit lets developers write, test, and deploy custom actions as organization tools. Custom code runs in an isolated runtime, and secrets stay separate from the code.
• Knowledge bases support website crawling with discovery preview, scheduled re-fetch, document editing, and multiple named retrievers with test-and-promote.
• The end-user experience was redesigned: streamed answers with clean partial rendering, generative interfaces such as forms, cards, charts, and date pickers, citations, and full keyboard and screen-reader support.
• Live chat is built in: a real-time inbox, assignment strategies, canned responses, CSAT, and a copilot for human agents. Human replies with rich content reach every active channel.
• WhatsApp connects through Meta Embedded Signup with template sync and campaign delivery from a prepaid campaign wallet.
• Analytics unifies agent, human, channel, and campaign results with drill-through to conversations and a build-your-own Explore view.
• Background work is unified: exports, imports, and bulk actions run in the background with visible progress in the task tracker.
• Security: a tamper-evident audit log with chain verification, SIEM streaming for Enterprise, bring-your-own-model-key, signed file downloads, plan-based data retention, and cold-storage archival.