Compendium

Chatbot Intent / NLU / Voice-Assistant DSLs Family Index

12 min read

Chatbot Intent / NLU / Voice-Assistant DSLs Family Index

type: language-family-index family: chatbot-intent-dsls languages_catalogued: 22 tags: [language-reference, family-index, chatbot-intent-dsls, nlu, voice-assistant, dialogflow, alexa, rasa, voicexml, scxml]

Chatbot Intent / NLU / Voice-Assistant DSLs — Family Index

Family overview

Chatbot intent / NLU DSLs are the textual languages used to declare the three artefacts of the pre-LLM conversational stack: (1) intents (the discrete user goals a bot can handle — BookFlight, OrderPizza, CheckBalance), (2) slots / entities (the parameters each intent needs — departure_city, pizza_size), and (3) dialog flow / fulfillment (what state the bot is in, what to ask next, when to call a backend). The reference architecture they all assume is the same: an utterance enters, a classifier picks the most likely intent and extracts entities, a dialog manager updates state, a fulfillment hook runs business logic, and an NLG layer (templates or response variants) emits a reply. This stack is roughly 30 years old — AIML 1.x (Wallace, ~1995–2002) is the recognisable ancestor — and was the orthodoxy from the smart-speaker boom (Alexa 2014, Cortana 2014, Google Assistant 2016) through the customer-service-bot wave of 2018–2022.

The 2023–2025 LLM disruption gutted this orthodoxy on a vendor-by-vendor schedule. Microsoft retired LUIS (Language Understanding Intelligent Service) on a phased path — portal access ended 31 October 2025, full retirement of authoring + inferencing APIs is 31 March 2026 — and pushed customers to CLU (Conversational Language Understanding), itself a thin transitional layer before Copilot Studio and Azure AI Foundry took over the dialog story. Twilio sunset Autopilot in August 2023, with no like-for-like replacement — Twilio’s official guidance was to bring your own Dialogflow CX or build on top of voice/SMS primitives. Amazon pivoted Alexa toward “Alexa+” / LLM-grounded experiences from 2023 onward (still rolling out as of 2026); Alexa Conversations remains technically active but deprecated in spirit (the dialog evaluation tool is already retired). Microsoft archived the Bot Framework Composer repository on 9 July 2025, putting Adaptive Dialogs JSON into formal read-only legacy status. Google kept Dialogflow CX alive but bolted Gemini-driven playbooks and generators onto it, blurring the line between intent flows and LLM agents.

What’s still alive in 2026 has clear reasons to be: VoiceXML survives in telephony / IVR because regulated, low-latency, deterministic call flows still rule out LLM-only stacks (despite the W3C Voice Browser WG closing in October 2015); Rasa survives in self-hosted / regulated industries that need data residency, with its CALM (Conversational AI with Language Models) engine bridging classical NLU and LLM dialog understanding; AIML survives because it deploys on a Raspberry Pi with no GPU and no API key; ChatScript survives in niche brand chatbots and voice-assistant authoring shops where Bruce Wilcox’s authoring expertise has been productised. The new wave — OpenAI Realtime API session config, Anthropic Claude tool-use schemas, and similar — replaces a hand-authored intent model + dialog tree with a few hundred tokens of system prompt and a tools[] array; the gpt-realtime GA announcement (2025) explicitly markets this as “voice agents without an interaction model.”

The arc is striking: 30 years of incrementally richer intent DSLs (AIML 1.0 → AIML 2.0 → ChatScript → RiveScript → LUIS → Dialogflow → Alexa Interaction Model → CLU → Adaptive Dialogs), each a careful schema for declaring the same three artefacts, are now mostly being replaced by a system prompt and a JSON tool schema. The classic intent DSL is becoming a niche compliance-and-determinism technology rather than the default conversational substrate.

In our deep library

None catalogued at deep-library tier. Cross-references:

  • ai-prompt-languages — the LLM-era successor family (Guidance, LMQL, DSPy, Outlines, prompt templates). The OpenAI Realtime API session config sits at the boundary and is genuinely dual-classified.
  • notation-spec — SCXML, CCXML, and VoiceXML are XML notation/spec languages and could equally live there.
  • visual-dataflow — Dialogflow CX, Bot Framework Composer, Botpress Studio, and Copilot Studio are visual dialog editors whose serialized form is the JSON/YAML DSL catalogued here.
  • python — host language for Rasa, RiveScript-Python, ChatScript bindings, the OpenAI Agents SDK.
  • javascript — host for Bot Framework SDK, Wit.ai SDKs, RiveScript-JS, OpenAI Realtime browser/WebRTC clients.
  • markup — AIML, VoiceXML, CCXML, SCXML are XML dialects and are dual-classified there.

Tier 3 family table

Language / DSLFirst appearedOriginVendor / PlatformStatus (2026)URL
AIML (Artificial Intelligence Markup Language)1.0 ~1995–2001; AIML 2.0 draft 2013Richard Wallace (A.L.I.C.E.) → AIML FoundationPandorabots (commercial host) + open Program AB / Program Y interpretersNiche but alive — AIML 2.0 is still the reference low-resource bot DSL; runs on RPi-class hardware with no GPU/API keyhttps://www.pandorabots.com/docs/aiml-reference/
ChatScript~2009Bruce Wilcox (4-time Loebner Prize winner: Suzette 2010, Rosette 2011)Open source (ChatScript/ChatScript on GitHub; older bwilcox-1234 repo abandoned)Active in niche brand-chatbot work; canonical repo migrated to org accounthttps://github.com/ChatScript/ChatScript
RiveScript2009 (Noah Petherbridge)Aichaos (open source)Independent OSS — JS, Python, Go, Java, Perl interpretersDormant / minimal maintenance — Snyk flags inactive; last sustained release activity 2022, sporadic sincehttps://www.rivescript.com/
Rasa NLU YAML training format1.0 (2019, replacing markdown format from Rasa 0.x)Rasa Technologies (Berlin)Rasa Open Source + Rasa ProActive — YAML format persists into the CALM era despite the dialog engine pivothttps://rasa.com/docs/rasa/training-data-format/
Rasa Stories2018Rasa TechnologiesRasa Open SourceLegacy in CALM — still parsed for backward compat; new bots use flows.ymlhttps://rasa.com/docs/rasa/stories/
Rasa RulesRasa 2.0, 2020Rasa TechnologiesRasa Open SourceLegacy in CALM; deterministic-override role partly subsumed by flow guardshttps://rasa.com/docs/rasa/rules/
Rasa Flows / CALM YAML2024 (Rasa Pro CALM GA)Rasa TechnologiesRasa Pro 2026Active — current authoring format; flow YAML + LLM-driven CommandGenerator replace stories/ruleshttps://rasa.com/docs/rasa-pro/concepts/flows/
Dialogflow ES Intent JSON2016 (as api.ai → Google Dialogflow)api.ai (acquired by Google 2016)Google CloudMaintenance mode — Google steers new work to CX; ES no longer receiving new featureshttps://cloud.google.com/dialogflow/es/docs/intents-overview
Dialogflow CX Flow JSON2020Google CloudGoogle Cloud (now part of Vertex AI Conversational Agents)Active, with 2024–2026 Gemini playbook + generator overlays blurring the intent-vs-LLM linehttps://cloud.google.com/dialogflow/cx/docs/reference/json-export
Microsoft LUIS .lu format2017Microsoft Azure Cognitive ServicesAzureRetiring 31 March 2026 — portal access ended 31 Oct 2025; APIs shut down 31 Mar 2026https://learn.microsoft.com/en-us/azure/ai-services/luis/whats-new
CLU (Conversational Language Understanding) JSON schema2021 (preview), GA 2022Microsoft Azure AI for LanguageAzure AI FoundryActive — the official LUIS migration targethttps://learn.microsoft.com/en-us/azure/ai-services/language-service/conversational-language-understanding/overview
Alexa Interaction Model JSON2014 (Alexa Skills Kit launch)AmazonAlexa Developer Console / ASK CLIActive but legacy-orientation — being eclipsed by Alexa+/LLM-grounded skill modelhttps://developer.amazon.com/en-US/docs/alexa/custom-skills/define-the-interaction-model-in-json-and-edit-it-with-a-text-editor.html
Alexa Conversations dialog model2020 (preview), GA 2021AmazonAlexa Skills KitDeprecated in spirit — dialog evaluation tool retired; Amazon’s strategic focus is Alexa+/LLMhttps://developer.amazon.com/en-US/docs/alexa/conversations/about-alexa-conversations.html
Watson Assistant skill JSON / dialog tree2016 (Watson Conversation) → 2018 renameIBMIBM Cloud (now watsonx Assistant)Active — watsonx Assistant integrates skill JSON with watsonx.ai LLM extensionshttps://cloud.ibm.com/docs/watson-assistant
Bot Framework Composer / Adaptive Dialogs JSON2019 (Composer preview), 2020 GAMicrosoftBot Framework v4 SDK + ComposerArchived 9 July 2025 — repo read-only; succeeded by Copilot Studiohttps://github.com/microsoft/BotFramework-Composer
Microsoft Copilot Studio topic YAML2023 (PVA), rebranded 15 November 2023Microsoft (Power Platform)Microsoft 365 / Power PlatformActive — Power Virtual Agents (2019) renamed Copilot Studio Nov 2023https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/microsoft-power-virtual-agents-now-part-of-microsoft-copilot-studio/
Twilio Autopilot intent model2018 (preview), 2019 GATwilioTwilioEnd-of-life August 2023 (end-of-support 25 Feb 2023) — no direct successor, customers ported to Dialogflow CX or build-your-ownhttps://support.twilio.com/hc/en-us/articles/8279929323547
Wit.ai schema (intents + entities + traits)2013 (founded), 2015 (Facebook acquisition)Wit.ai → MetaMeta for Developers (free NLP service)Active but de-prioritised — Meta’s developer messaging now emphasises Llama / generative routeshttps://wit.ai/docs
Botpress NLU intent / utterance config2017 (Botpress v1)Botpress (Sylvain Perron, Quebec)Open core + Botpress CloudActive — Botpress Studio is now AI-Transition-Card-first; legacy Library tab deprecated, classic NLU still supportedhttps://botpress.com/docs
OpenAI Realtime API session configPublic beta Oct 2024; gpt-realtime GA late 2025OpenAIOpenAI APIActive and ascendant — the de facto modern voice-agent “interaction model”https://platform.openai.com/docs/guides/realtime
VoiceXML 2.0 / 2.1VXML 2.0: W3C Rec 16 Mar 2004; VXML 2.1: W3C Rec 19 Jun 2007W3C Voice Browser WG (closed Oct 2015)Vendor IVR platforms (Cisco CVP, Nuance, Genesys, Voxeo/Aspect, Voicent, etc.)Surviving standard — entrenched in telephony/IVR; VXML 3.0 stalled at WD; some vendor deprecations starting (Cisco CVP gateway, late 2025)https://www.w3.org/TR/voicexml21/
CCXML (Call Control XML)W3C Rec 5 Jul 2011W3C Voice Browser WGIVR platform vendors (paired with VoiceXML)Surviving standard — narrow scope, used wherever VoiceXML ishttps://www.w3.org/TR/ccxml/
SCXML (State Chart XML)W3C Rec 1 Sep 2015W3CUnderlies many dialog managers (Apache Commons SCXML, JSSCXML, etc.)Active — adopted as a generalised state-machine substrate beneath modern dialog systemshttps://www.w3.org/TR/scxml/

Notable threads

  • 30 years of incrementally richer intent DSLs vs. ~3 years of LLM disruption. AIML 1.0 (~1995) and AIML 2.0 (draft 2013) staked out pattern-and-template authoring; ChatScript (~2009) added concepts and topic stacks; RiveScript (2009) simplified the syntax; LUIS (2017), Dialogflow (2016/2020), Alexa Interaction Model (2014), Wit.ai (2015), and Watson Assistant (2016) commercialised cloud-hosted intent classification at scale. Each generation was a more powerful schema for declaring the same three artefacts (intents, slots, dialog). Then 2023 happened: GPT-4 + tool use turned a 2,000-line intent JSON into a system prompt and a tools[] array. The DSL doesn’t do less — it just doesn’t need to be the user-facing artefact anymore.

  • The LUIS deprecation story is a textbook platform-commitment lesson. Microsoft announced LUIS retirement in 2022 with a 30 Sep 2025 target, then extended select APIs to 31 March 2026 in response to migration friction. Portal access ended 31 October 2025. The successor (CLU) is itself transitional — Microsoft’s strategic centre of gravity is now Copilot Studio (post-15 Nov 2023 rebrand) and Azure AI Foundry. Customers who built deeply on LUIS faced a migration path that pointed at a moving target; CLU may itself be subsumed into Foundry intent-detection primitives within a few years. Net lesson: cloud NLU DSLs are tightly coupled to vendor strategy cycles and outlive a typical 5-year enterprise app horizon poorly.

  • Alexa Conversations (2020–) as the brief ML-bridge era. Between hand-authored interaction models (Alexa 2014–2019) and pure LLM (Alexa+, 2023+), Amazon shipped Alexa Conversations: an ML-driven dialog manager that learned dialog policy from annotated training data instead of explicit dialog trees. The concept — let an ML model learn dialog flow from examples — turned out to be a half-step that LLMs leapfrogged within four years. The dialog evaluation tool is already retired per the ASK deprecated-features page; the feature itself is technically still callable but is not where Amazon is investing.

  • VoiceXML as the survivor. Telephony / IVR systems still author dialog in VoiceXML 2.0/2.1 in 2026, despite the W3C Voice Browser Working Group having closed in October 2015 and VoiceXML 3.0 never advancing past Working Draft. Two reasons: (1) regulatory — financial-services and healthcare IVR has audit, retention, and explainability requirements that LLM-only stacks can’t yet satisfy; (2) reliability and latency — telephony has hard SLAs that don’t tolerate LLM tail-latency or hallucinated responses on a billing call. There are signs of erosion (Cisco starting CVP-gateway deprecations in late 2025) and AI-IVR vendors (DataKnowl DML, etc.) are pitching VoiceXML alternatives, but the installed base is large and slow-moving.

  • Rasa’s bet on open-source NLU + dialog still works in 2026. While the major cloud platforms pivoted to LLM-centric stacks, Rasa’s CALM (Conversational AI with Language Models, GA 2024) introduced a hybrid where flow YAML provides the deterministic skeleton and an LLM-based CommandGenerator handles dialog understanding. The YAML training-data format from Rasa 1.x is still the substrate. Rasa Pro 2026 holds onto the regulated-industry / data-residency / self-hosted niche — banks, telcos, governments that can’t or won’t ship customer utterances to OpenAI/Anthropic.

  • The OpenAI Realtime API session config as the new “interaction model” — but smaller. A traditional Alexa skill might ship a 3,000-line Interaction Model JSON declaring 50 intents, 200 slots, and dialog directives. The equivalent voice agent on the Realtime API is a session.update event with a system prompt, a tools[] array, and a voice selection (alloy, cedar, marin, etc.). The DSL hasn’t disappeared; it’s collapsed by an order of magnitude because the LLM does intent classification, slot filling, and NLG implicitly. The 2025 gpt-realtime GA + remote MCP support + SIP phone calling means even the telephony niche has a credible LLM-native challenger — though the determinism gap remains.

  • AIML in 2026: the zero-resource floor. AIML survives precisely because everything that replaced it requires a GPU, a cloud account, or an API key. A Pandorabots-hosted AIML brain (or a self-hosted Program AB / Program Y) runs on commodity hardware, has predictable latency, and the category-template model is debuggable by non-engineers. Use cases: low-bandwidth deployments (rural FAQ kiosks, embedded toys, museum exhibits), training/teaching contexts, and brand chatbots where the long tail of questions is small and known.

  • The “JSON or YAML?” split tracks vendor culture. Microsoft (LUIS .lu, CLU JSON, Adaptive Dialogs JSON, Copilot Studio topic YAML) and Google (Dialogflow ES + CX both JSON) lean JSON; Rasa is YAML-first across the board; Amazon’s Alexa Interaction Model is JSON; Watson Assistant is JSON; AIML/VoiceXML/CCXML/SCXML are XML by W3C heritage. There’s no technical reason for any of these to be one or the other — it’s a cultural/tooling choice that affects who can author them (data engineers tend YAML, web devs tend JSON, enterprise integration shops tend XML).

Citations

Graph View