Guides
Guides
Breadcrumbs

November 2026


Agentic Inference Server 1.1.0

New features

  • MLFlow support for main model

  • Drafter/speculative decoding integration, with MLFlow support for drafter model

  • Auth/authz implementation

Improvements

  • Expanded hardware/runtime compatibility, and adjusted per-hardware configs

  • Overall throughput improvement, which reduces cost

    • ~0.2 req/s throughput improvement for Q&A (0.6->0.8 for bs=1, 1.2->1.5 for bs=5

    • 0.6-3 req/s throughput improvement for Router (2.7->3.3 for bs=1, 20.5->23.4 for bs=20)

  • Overall latency improvement, which improves CX (reported numbers for autoregressive; expect even lower ITL with speculative decoding)

    • 1100->900ms P50 TTFS, 12.6->9.5 avg ITL for Q&A cc=1

    • 620->560ms P50 TTFS, 26.4->21.3 avg ITL for Router cc=20

AudioHub 7.5.0

Improvements

  • QoS: During a speech-synthesis outage, calls now switch straight to the backup voice instead of waiting before every prompt. The standard voice is restored automatically once the service recovers.

Security

  • General security vulnerability fixes.

Copilot 1.20.0

Improvements

  • Platform upgrade — The Copilot interface moves to a newer UI framework, which keeps it secure and up to date. There are no visible changes.

This release includes security updates.

Liveness Detection 7.3.0

RA Detection model — v4.0.5

  • The current trend is to use bots to complete users' tasks. Such bots may use cheaper standard TTS (concatenative synthesis). In order to improve detection of these bots, we intend to enlarge our training set for liveness detection models with relevant data.

Security

  • Resolved security vulnerabilities

OXP 1.23.0

Fixed OCP xPacks

  • en-US Universal deepNLU® xPack v3.22.0 / en-US Universal+ deepNLU® xPack v3.22.0-Addresses

    • Tuning for the recognition of the following entities: Alphanumeric, Alpha, CardNumber, CAZipCode, Date, DIALOGACT, SentimentDescriptor, and Text.

    • Added recognition of spoken references to letter case (e.g. "capital letter", "lower case") and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • A typed "-" in Alphanumeric input is now handled like the spoken "dash": a value such as "a-b-c" is returned as one value instead of being split.

    • Typed Alphanumeric input of two to four space-separated groups of letters and digits (e.g. "A123 4567", "As 1234") is now returned as one value of up to 27 characters instead of being cut.

    • Alphanumeric now accepts an all-digit code of 22 to 27 digits, and a typed letter code whose digit run is 22 to 27 digits, as one value.

    • Improved recognition of long spoken card numbers (CardNumber) that contain "hundred".

    • Fixed an 8-digit spoken date month-first containing a 0 (e.g. "zero one one five two zero zero eight") returning an impossible Year. It now resolves to the correct day, month and year.

    • Fixed "twenty thirty three" being resolved to the year 203 instead of 2033.

    • "nineteen oh one" to "nineteen oh nine" now return Year=1901 to 1909 instead of 1900, and "nineteen hundred" and "one thousand nine hundred" now return Year=1900.

    • "fortnight" is now read as two weeks on the date prompts (e.g. "in a fortnight" returns TimeRelativeToNow=2 and "a fortnight ago" returns TimeRelativeToNow=-2).

    • Fixed "hang on" and "hang tight" returning a negative SentimentDescriptor.

    • Improved Alphanumeric recognition of a doubled letter written as one word (e.g. "xx") when an alphanumeric is requested.

    • A spelled "aa" or "ee" after a letter in Alphanumeric input (e.g. "pb aa nine nine ...") is now kept instead of being dropped as a hesitation.

    • A Canadian postal code typed in capitals with a space (CAZipCode) is now returned as a single Alphanumeric value with the same code, instead of one value per half.

    • Fixed Alpha input containing "vertical line" returning an extra LocationIndicator and USCity.

    • DIALOGACT=CLOSING is now returned for "okay thank you" after the re-ask closing prompt (Intent_ReAsk).

    • DIALOGACT=CLOSING is now returned for "not at this time" and "not at the moment" on the anything else prompt (Intent_ReAsk), as for "not today".

    • "i don't have it" no longer returns CONVENTIONAL=notAvailable twice.

    • "i don't recall" (also "can't recall" and "do not recall") now returns CONVENTIONAL=notAvailable and MetaAction=info, the same as "i don't remember". A plain "recall" on the Hold_MaxWait prompt still returns CONVENTIONAL=callBack.

    • Text miniApp confirmation step: a plain "yes" or "no" no longer returns Text. A correction such as "no i said john smith" still does.

  • en-US Universal 2.0 deepNLU® xPack v2.9.0 / en-US Universal 2.0+ deepNLU® xPack v2.9.0-Addresses

    • Updated the dependencies from en-US Universal v3.21.0 to v3.22.0. Includes all en-US Universal fixes and improvements released in between.

  • en-US Banking 2.0 deepNLU® xPack v4.7.0

    • Updated the dependencies from en-US Universal v3.21.0 to v3.22.0. Includes all en-US Universal fixes and improvements released in between.

  • en-US Energy deepNLU® xPack v3.4.0

    • Updated the dependencies from en-US Universal v3.21.0 to v3.22.0. Includes all en-US Universal fixes and improvements released in between.

    • Fixed the App entity not being created for utterances such as "i have trouble with gmail on my phone" and "trouble with yahoo on my tablet" (gmail, yahoo, yahoo mail and google).

  • en-US Car Retail deepNLU® xPack v3.4.0

    • Updated the dependencies from en-US Universal v3.21.0 to v3.22.0. Includes all en-US Universal fixes and improvements released in between.

  • en-GB Universal deepNLU® xPack v3.19.0 / en-GB Universal+ deepNLU® xPack v3.19.0-Addresses

    • Tuning for the recognition of the following entities: Alphanumeric, CardNumber, Date, DIALOGACT, SentimentDescriptor, and Text.

    • Added recognition of spoken references to letter case (e.g. "capital letter", "lower case") and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • A typed "-" in Alphanumeric input is now handled like the spoken "dash": a value such as "a-b-c" is returned as one value instead of being split.

    • Typed Alphanumeric input of two to four space-separated groups of letters and digits (e.g. "A123 4567", "As 1234") is now returned as one value of up to 27 characters instead of being cut.

    • Alphanumeric now accepts an all-digit code of 22 to 27 digits, and a typed letter code whose digit run is 22 to 27 digits, as one value.

    • Improved recognition of long spoken card numbers (CardNumber) that contain "hundred".

    • Fixed "twenty thirty three" being resolved to the year 203 instead of 2033.

    • "nineteen hundred" and "one thousand nine hundred" now return Year=1900.

    • "fortnight" is now read as two weeks on the date prompts (e.g. "in a fortnight" returns TimeRelativeToNow=2 and "a fortnight ago" returns TimeRelativeToNow=-2).

    • Date miniApp: date answers such as "last week" or "in two weeks" no longer return FirstName or LastName.

    • Fixed "hang on" and "hang tight" returning a negative SentimentDescriptor.

    • Improved Alphanumeric recognition of a doubled letter written as one word (e.g. "xx") when an alphanumeric is requested.

    • A spelled "aa" or "ee" after a letter in Alphanumeric input (e.g. "pb aa nine nine ...") is now kept instead of being dropped as a hesitation.

    • DIALOGACT=CLOSING is now returned for "okay thank you" after the re-ask closing prompt (Intent_ReAsk).

    • DIALOGACT=CLOSING is now returned for "not at this time" and "not at the moment" on the anything else prompt (Intent_ReAsk), as for "not today".

    • "i don't have it" no longer returns CONVENTIONAL=notAvailable twice.

    • "i don't recall" (also "can't recall" and "do not recall") now returns CONVENTIONAL=notAvailable and MetaAction=info, the same as "i don't remember".

    • Text miniApp confirmation step: a plain "yes" or "no" no longer returns Text.

    • Confirmed that Numeric accepts up to 21 digits and Alphanumeric up to 27 characters.

  • en-AU Universal deepNLU® xPack v3.10.0

    • Tuning for the recognition of the following entities: Alphanumeric, CardNumber, Date, DIALOGACT, SentimentDescriptor, and Text.

    • Added recognition of spoken references to letter case (e.g. "capital letter", "lower case") and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • A typed "-" in Alphanumeric input is now handled like the spoken "dash": a value such as "a-b-c" is returned as one value instead of being split.

    • Typed Alphanumeric input of two to four space-separated groups of letters and digits (e.g. "A123 4567", "As 1234") is now returned as one value of up to 27 characters instead of being cut.

    • Alphanumeric now accepts an all-digit code of 22 to 27 digits, and a typed letter code whose digit run is 22 to 27 digits, as one value.

    • Fixed day-first dates with a zero-led day, month 11 or 12 and a spoken year (e.g. "zero two one one nineteen sixty five") returning day and month swapped. They now resolve day-first (DayOfMonth=2, Month=11, Year=1965).

    • Improved recognition of long spoken card numbers (CardNumber) that contain "hundred".

    • Fixed "twenty thirty three" being resolved to the year 203 instead of 2033.

    • "nineteen oh one" to "nineteen oh nine" now return Year=1901 to 1909 instead of 1900, and "nineteen hundred" and "one thousand nine hundred" now return Year=1900.

    • "last fortnight" now returns TimeRelativeToNow=-2 instead of 2, and "a fortnight from today" returns TimeRelativeToNow=2 alone.

    • Fixed "hang on" and "hang tight" returning a negative SentimentDescriptor.

    • Improved Alphanumeric recognition of a doubled letter written as one word (e.g. "xx") when an alphanumeric is requested.

    • A spelled "aa" or "ee" after a letter in Alphanumeric input (e.g. "pb aa nine nine ...") is now kept instead of being dropped as a hesitation.

    • DIALOGACT=CLOSING is now returned for "okay thank you" after the re-ask closing prompt (Intent_ReAsk).

    • DIALOGACT=CLOSING is now returned for "not at this time" and "not at the moment" on the anything else prompt (Intent_ReAsk), as for "not today".

    • "i don't have it" no longer returns CONVENTIONAL=notAvailable twice.

    • "i don't recall" (also "can't recall" and "do not recall") now returns CONVENTIONAL=notAvailable and MetaAction=info, the same as "i don't remember".

    • Text miniApp confirmation step: a plain "yes" or "no" no longer returns Text.

  • fr-FR Universal deepNLU® xPack v3.4.0 / fr-FR Universal+ deepNLU® xPack v3.4.0-Addresses

    • Tuning for the recognition of the following entities: Alphanumeric, DIALOGACT, and SentimentDescriptor.

    • Added recognition of spoken references to letter case and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • Fixed a typed Alphanumeric code containing a single capital letter (e.g. "A12345678") losing characters. The code is now returned whole.

    • Resolved a conflict between AGENT_REQUEST and CLOSING/NEW_TASK values. When an utterance contains AGENT_REQUEST together with CLOSING or NEW_TASK, only AGENT_REQUEST is now returned.

    • Confirmed SentimentDescriptor entity support; the detected polarity is returned as the value of the entity.

    • Removed "@" prefixes from dictionary tokens to avoid warnings during contextualization on OCP.

  • fr-CA Universal deepNLU® xPack v3.12.0

    • Tuning for the recognition of the following entities: Alphanumeric, CAZipCode, DIALOGACT, and SentimentDescriptor.

    • Added recognition of spoken references to letter case and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • Fixed a typed Alphanumeric code containing a single capital letter (e.g. "A12345678") losing characters. The code is now returned whole.

    • A Canadian postal code recognized as CAZipCode (e.g. "E1A 1N8") now always comes with one Alphanumeric value with the same code. Previously a letter could be lost.

    • Resolved a conflict between AGENT_REQUEST and CLOSING/NEW_TASK values. When an utterance contains AGENT_REQUEST together with CLOSING or NEW_TASK, only AGENT_REQUEST is now returned.

    • Confirmed SentimentDescriptor entity support; the detected polarity is returned as the value of the entity.

    • Removed "@" prefixes from dictionary tokens to avoid warnings during contextualization on OCP.

  • fr-BE Universal deepNLU® xPack v3.2.0

    • Tuning for the recognition of the following entities: Alphanumeric, DIALOGACT, and SentimentDescriptor.

    • Added recognition of spoken references to letter case and of special characters commonly used in passwords in Alphanumeric input. They are recognized but not included in the resulting value.

    • Fixed a typed Alphanumeric code containing a single capital letter (e.g. "A12345678") losing characters. The code is now returned whole.

    • Resolved a conflict between AGENT_REQUEST and CLOSING/NEW_TASK values. When an utterance contains AGENT_REQUEST together with CLOSING or NEW_TASK, only AGENT_REQUEST is now returned.

    • Confirmed SentimentDescriptor entity support; the detected polarity is returned as the value of the entity.

    • Removed "@" prefixes from dictionary tokens to avoid warnings during contextualization on OCP.

Pathfinder 1.9.8

New Features

  • Knowledge Base Search Index for Partial or Misspelled Names — Large document-based knowledge bases, such as a list of hotels, can now be searched by a partial or misspelled name instead of requiring an exact match. A new "Search index" view lets you upload a list of names and details, track processing status, test a search, and remove the index when it's no longer needed. The underlying search returns ranked matches with a confidence score, asking for more details only when several entries match equally well.

  • Adjustable Match Strictness for Knowledge Base Search — Knowledge base projects can now set their own match strictness — Strict, Balanced, or Lenient — directly in Project Settings. This lets you decide whether search results must match closely or return broader context, without needing to tune it on every request.

Fixes

  • More Reliable Conversations During Multi-Turn Search — When the AI service that rephrases follow-up questions is slow or unavailable, search now automatically falls back to your original question instead of stalling, keeping multi-turn conversations responsive.

  • Fewer Failed Document Uploads — Documents and web pages occasionally failed to process because of temporary connection issues reaching the encoding service. These are now automatically retried, so more uploads succeed without manual re-processing.

  • Correct Inline Display for Certain PDFs — Some PDF documents in the knowledge base viewer downloaded instead of opening directly in the browser. They now display inline as expected.

  • Accurate Search After Rolling Back an FAQ Document — Rolling back a frequently-asked-question document to an earlier version used to make its answer disappear from search results. Search now continues returning the correct answer after a rollback.

Self Learning 1.21.0

New Features

  • Word boosting — A new Word Boosting section in Settings lets you preview the words each agent should recognise better, then review and accept them. The accepted lists are applied when you deploy.

  • Agent testing — Testing has a new Agents tab for unit-testing concierge, routing and Q&A agents, and all test runs are shown in one combined Runs list.

Improvements

  • Long test runs keep going — Test runs now use their own service credentials, so long runs no longer stop when your sign-in session expires.

  • Drop files anywhere — Files can be dropped anywhere on the page, even while you're typing in a field.

  • Focused QM reviews — A Quality Management review scoped to specific agents now uses only those agents' calls.

Fixes

  • Security fixes in the web interface.

This release includes security updates.

Self Learning 1.21.2

Fixes

  • Voices load for projects not yet deployed — The persona voice list now loads for projects that haven't been deployed yet.

Self Learning 1.21.3

New Features

  • Escalation tests — Concierge tests can now check that conversations escalate to a human queue when they should, and the Agents tab in Testing shows the escalation queues.

Self Learning 1.21.4

Improvements

  • Voice list order — Lexis voices are now listed first, followed by Amazon Polly voices.

Self Learning 1.21.5

Fixes

  • Faster, more reliable call imports — Calls imported from your contact centre are no longer downloaded more than once, and imports no longer slow down as the backlog grows.

  • Plain-text transcripts saved as .json — Transcript files named .json that contain plain text are now imported as text instead of staying pending. Files that can't be read are marked as failed instead of staying pending, and a corrected file with the same name can be uploaded again.

This release includes security updates.

Self Learning 1.22.0

New Features

  • Backfill that never stalls — Recordings that cannot be studied (missing audio, failed imports) no longer block a project's initial history study; stuck projects recover on their own.

  • Controlled cost and trace visibility — Usage, cost and stored-call-trace views are restricted to verified platform staff, with a separate allowlist for raw call traces.

  • Gentler dashboards — Behind a flag: Mission Control boards use caching and bounded queries, so a slow database cannot be overloaded by dashboard refreshes.

Fixes

  • Security fixes. Plain-text transcripts named .json are imported correctly (already shipped with 1.21.5).

This release includes security updates.

Voice Biometrics 1.18.2

This is a targeted hotfix patch release for BioStore for Frankfurt environment.