Press CTRL+K to open search

Version 5.10

  • Published on Oct 19, 2026
  • 10 minute(s) read
  • Suara Atay
 Prev Next

AI Agents Version 5.10

19 October, 2026

AI Agents

  • Flow Builder with AI:
    Flows can now be built by describing them in plain language, so a subflow or a new branch takes one conversation instead of placing nodes one by one. The "Build with AI" button in the flow designer toolbar opens a panel on the right, and the canvas updates live as the assistant adds, changes or removes nodes. It works on Draft versions only, needs its LLM settings on the Flow Builder tab in Tenant Settings, and closing the panel starts a fresh conversation.

  • Knowledge Base:

    • Viewing Knowledge Base Chunks:
      You can now see the chunks each document or website page was split into, so you can check what your chunking settings produce and trace retrieval problems to their source. A "View chunks" action on each Indexed document in Data Sources, and on each row of the Indexed URLs page, opens a Chunks page you can search. Chunks are read-only and always show the source as it was last processed, in both draft and published versions.
    • Tag Parameters for the Knowledge Retrieval Tool:
      An AI agent can now narrow knowledge retrieval by tag on its own, picking the tag value that fits the user's question instead of always searching the whole knowledge base. "Add Tag Parameter" in the Knowledge Retrieval tool's Parameters section turns a knowledge base tag into an optional parameter, and you can edit its description and "Predefined Options" (up to 100 values). Tag names can now contain only letters, numbers and underscores, and a tag used by an AI agent tool cannot be edited or deleted.
    • Tags in the Knowledge Base Report:
      The Knowledge Base Responses page under Insights now shows the tags a retrieval was narrowed by and the tags on every retrieved chunk, so you can tell whether an empty result came from the filter or from the content. An "Applied Tags" block appears under the user query, each chunk shows its own tags, Not Found records now open the detail panel, and a call that returned nothing shows "No chunks found." Records created before this release show no tags.

  • AI Insights:

    • Separate Settings for AI Insights and Custom Attributes:
      AI Insights and Custom Attributes can now be turned on separately, so you can use Custom Attributes without setting up an LLM. Each has its own toggle on a new Conversation Attributes tab in Tenant Features and in Project Settings > General, and AI Insights now takes its LLM from a new page under Administration > Settings > AI Insights. After the upgrade, existing tenants and projects keep their current state, and a draft cannot be published while AI Insights is on but its LLM configuration is incomplete.
    • Generating Insights for Past Conversations:
      AI Insights can now be generated for past conversations, so conversations that missed generation, or that happened before the feature was turned on, are no longer left out of analysis. Set a "Generate Insights From" date on the Administration > Settings > AI Insights page, and completed conversations from that date onward that have no insights are processed in the background. Before saving an earlier date, you are asked to confirm and shown how many conversations will be processed, because each one uses LLM capacity and this cannot be undone.
    • Output Language for AI Insights:
      Free-text insights such as the summary are now written in a language you choose for the project, instead of one guessed from the conversation. Set it in the "Output Language" field on the AI Insights tab of Project Settings > Conversation Attributes; it starts on the project's primary language. Category values stay as defined, and insights generated before a change keep their original language.

  • Intent Recognition Report:
    A new Intent Recognitions page under Insights now lists every recognised utterance with the NLU's top intent, its confidence and the final intent, so you can investigate a single misrouted turn instead of only seeing totals on the dashboard. When Advanced Conversation Routing is on, each row also shows its outcome, such as Confirmed, Overridden or Rescued, and the detail panel lists the candidate intents with their scores. The page has its own permission, and no role has it by default.

  • Integrations:

    • Stored Credentials for HTTP Requests:
      You can now store tokens and secrets for HTTP requests once and reuse them, instead of typing them into each HTTP node's headers in plain text. A new Credentials page under Integrations holds encrypted Custom Header and OAuth 2.0 Client Credentials entries, and the HTTP node's new "Credential" field picks one, fetching the access token automatically for OAuth 2.0. An OAuth 2.0 credential is saved only once a token can be obtained, and a credential used by a draft or published flow cannot be deleted.
    • Client Certificates and mTLS:
      HTTP nodes can now call services that require mutual TLS by presenting a client certificate stored on the platform. Upload certificates as PFX or PEM files, with an optional CA certificate, on a new Certificates page under Integrations, then use the "Enable mTLS" toggle and "Client Certificate" field in the HTTP node's Advanced Settings. An updated certificate file can take up to 24 hours to reach live conversations, expired certificates cannot be selected, and an exported flow drops its certificate selection, so you need to choose it again after import.

  • AI Testing:

    • Correct Resolution Criterion and Evaluation Reasons:
      A "Correct Resolution" template has been added to the success criteria templates, scoring whether the AI agent actually resolved the customer's request rather than only how well it spoke. Evaluation reasons now appear on a Success Criteria tab in the Conversation panel of Evaluation Details, or when you click a criterion cell, instead of in a hover tooltip. Criterion descriptions can now be up to 2,000 characters long and keep their line breaks.
    • Tags for AI Tests:
      You can now tag AI tests, so problematic simulations are easy to find again and share with the right team. The "Tags" field on the test definition, for simulated conversations and historical data alike, takes up to 5 tags of up to 50 characters each. A tag filter next to the search box on the AI Testing page shows only tests that carry at least one of the selected tags.
    • Unscored Conversations in Test Results:
      Test run results now show how many conversations were actually scored, so you notice when Overall Performance is based on fewer conversations than you expected. The run detail shows "Scored Conversations" as a count out of the total, shows a warning when some could not be scored, and marks those conversations in the session list. The Overall Performance calculation itself is unchanged.

  • End-to-End Call Recording for IVR Conversations:
    You can now play back an IVR conversation as one end-to-end recording of the whole call, including silences, hold, announcements and transfers, when the Data Flow voice gateway delivers it through a new external API endpoint. The "Audio Player" setting under the gear icon on the Conversations page switches between "Per Message" and "Entire Call Audio", and the play button pauses and resumes the recording. Conversations without a full recording play message by message as before.

  • AI Agent Invocation Limits:
    You can now cap how many times AI agents are called within one user turn and within a whole session. This stops runaway hand-offs between agents that keep a customer waiting and multiply LLM cost. The "AI Agent Invocation Limit" toggle and its "Max Invocations Per Turn" and "Max Invocations Per Session" fields are on a new Project Limitations tab in Project Settings. The limit is off by default, and once it is reached the flow continues from the AI Agent node's Failed output.

  • Separate Permission for Exporting Conversations:
    You can now let users review conversations without letting them export them, for teams that need to limit who takes personal data off the platform. A new "Export" permission under Conversations in the Roles page's permissions window controls the Export buttons on the conversation list and the conversation detail. After the upgrade, roles that can see conversations keep the right to export them, while new roles start without it.

Improvements

AI Agents

  • Knowledge Base:

    • Read-Only Knowledge Base View:
      Users who can view projects but not edit them can now open a knowledge base read-only, instead of seeing only the list. They can see its settings, data sources, indexed URLs, chunks and tags, and download documents, while every editing action stays hidden.
    • Knowledge Base Tags in RAG Agent Responses:
      The RAG Agent node's response now includes the knowledge base tags of each returned document, as the Knowledge Retrieval node's already does, so flows can route on them or report where an answer came from. The response body sample in the node panel shows the new field.

  • Automatic Delivery of BotHub Settings:
    BotHub settings now reach BotHub automatically after deployment and whenever they change, so nobody has to open the page and press Save on a new or upgraded environment. The Administration > Settings > BotHub Settings page shows whether the settings have been delivered and when, and a new "Renew credentials" button generates a new client secret and sends it to BotHub. The four client ID and secret fields are now generated by the platform and read-only.

  • License Credit Consumption by Resolution Type and Duration:
    License credits can now be consumed according to how a session ended, using Resolution Type coefficients defined on the license, so billing reflects sessions that ended without customer interaction or with a live agent handover. Each consumption appears in the License Portal labelled with the session's Resolution Type. Minute-based licenses consume the session's duration, rounded up to the license's interval before the coefficient applies, and licenses without coefficients still consume one credit per session.

  • Reasoning Effort for Azure OpenAI Configurations:
    You can now set a reasoning effort level for Azure OpenAI configurations, as you already could for custom models, on deployments that support reasoning. Wherever you select an Azure OpenAI configuration, turning on the "Enable" toggle under "Reasoning Effort" lets you choose Default, None, Low, Medium or High. The platform cannot check whether a deployment supports reasoning, so if it does not, the provider returns an error during conversations.

  • Listing Projects Through the External API:
    You can now list a tenant's projects through the external API, so integrating products can find out which projects exist without manual setup. The call uses the tenant's existing client credentials token, takes no parameters, and returns each project's ID (the same ID used in conversation reports), name, description, publish state, published version, and creation and modification times. Each IP address can make up to 5 requests per second.

  • Turning Off Log Shipping and Trace Export:
    On-premises deployments can now switch off log shipping to Elasticsearch and trace export to an APM server, each on its own, through deployment configuration. With either one off, services start without that address configured, while console and file logs and Prometheus metrics are unaffected. When log shipping is off, the Traces tab on the conversation detail and the Log tab in Test Chat are hidden; both stay on unless you turn them off.

  • Live Agent ID in the Conversational Analytics Integration:
    After a handover, conversations sent to Conversational Analytics now carry the live agent's own ID, so they map to the right agent there. When no live agent ID is available, the existing identifier is sent as before.

  • Interrupted Interim Messages No Longer Repeated:
    If a caller talks over the interim message an AI agent plays while a tool runs, the agent no longer repeats that message once the tool returns. Interruptions of any other answer are handled as before, and the interrupted message still appears in the transcript and reports.

  • No AI Insights for AI Testing Conversations:
    AI Insights are no longer generated for AI Testing simulation conversations, even when the project has AI Insights turned on, which avoids LLM cost on test traffic.

  • Minor UI Improvements:
    Several minor improvements have been made to the user interface to
    enhance user experience and usability.

  • Minor Bug Fixes:
    Various minor bugs have been addressed to improve overall system
    stability and performance.