IBM's "6 types of chatbots" explainer (dated 7 March 2025), Tidio's "Types of Chatbots" article (updated 2 July 2026) and Tidio's help article "How do flows, Lyro, and live agents work together?" (dated 4 September 2025), Google's Dialogflow CX pages "Generative versus deterministic" and "General agent design best practices" (both stamped 3 September 2026 UTC) and "Data store hybrid agents" (stamped 29 July 2026 UTC when fetched and 3 September 2026 UTC when re-read an hour later, identical text; both recorded in Sources), the Rasa "CALM Overview" page (last updated 10 September 2026), Botpress's "Autonomous Node" reference, Manychat's "AI Step" article (updated 31 July 2026), Landbot's "AI Agent Chatbots" product page and its "Is Landbot a Chatbot or an AI Agent?" post (20 February 2026), Tidio's "Disable Text Input" article (dated 15 May 2025) and Voiceflow's "Buttons" step documentation were read on 11 September 2026, and every quotation and priority order below carries that date; vendors change these pages without notice. Every count of our own reviews is a tally of the fifteen files matched by sample-reviews/*-review.md, with the search string printed in Sources.

Hybrid chatbot· Chatbot architecture

A hybrid chatbot is a bot that runs two kinds of logic under one conversation: scripted flows (buttons, menus, if-then rules, forms) for the paths its builder can predict, and a language model or NLU layer for the messages they cannot. IBM's definition is a system that combines rule-based logic with machine learning capabilities. Vendors use the same word for two other pairings, a bot that hands the conversation to a human agent, and one bot that answers on both voice and chat, so the useful question is which layer answers first when a message arrives. In the flows-plus-model sense, that order is a documented setting: Tidio runs flows, then Lyro tasks, then the Lyro AI agent, then people; Dialogflow CX evaluates form parameters, then intent routes, then its data stores.

By Chatbotscape Editorial· Methodology· Published 11 September 2026· Updated 11 September 2026

Hybrid Chatbot — The Three Things the Word Means, and Which One You Are Being Sold

Quick answer: A hybrid chatbot pairs a scripted layer with a model layer. The scripted layer is the flow builder: buttons, menus, keyword triggers, forms that collect a name and an order number in a fixed sequence. The model layer is whatever reads a typed sentence the flow did not anticipate: an intent classifier in the older products, a large language model with a knowledge base in most of the ones we review now. IBM's explainer, read 11 September 2026, puts it in one sentence: "A hybrid chatbot is a conversational AI system that combines rule-based logic with machine learning capabilities." Tidio's own article on chatbot types says the same thing from a vendor's side, "Hybrid chatbots combine a rule-based layer with an AI layer, so they can run fast scripted flows and still handle open-ended questions," and then adds two more meanings later in the same article: hybrids "also pair the bot with live chat," and voice bots "usually split into hybrid bots (text and voice) and voice-only bots." Three architectures, one word.

The word is worth pinning down because the products have converged on it. Of the fifteen platforms we review, one (Landbot) sells itself with the phrase "the control of rule-based chatbots"; the phrase "flow builder" appears in thirteen of the fifteen review files and "knowledge base" in all fifteen (string counts, printed in Sources), and the products that have both ship them in the same login without calling the result anything. What separates them is not whether they are hybrid but how the seam works: which layer gets the message first, whether the model can interrupt a flow, and what happens to the data when one layer hands to the other. This entry defines the term, separates the three senses, and reads the seam rules the vendors actually document.

What this entry is and is not

Definitions, priority orders and quotations are read from the vendor and reference pages named in Sources, each dated. Tidio, Landbot, Manychat, Botpress, Voiceflow, AiSensy and Blip have Chatbotscape reviews that carry affiliate links; IBM, Google and Rasa do not; see our affiliate disclosure. This entry owns the word: its three senses, and how the flows-plus-model seam is ordered on the platforms that document it. The four-architecture taxonomy (rule-based, NLU intent classification, generative, hybrid) lives in what is a chatbot, which describes the hybrid as three layers, rule flows plus NLU plus an LLM; this entry treats the classifier and the LLM as one layer, the one that reads typed text, because the seam question is the same for both; the flow-by-flow decision of what to move from rules to a model is the generative vs rule-based guide; the per-question choice between a button and a typed answer, which is what a hybrid is built out of, is the buttons vs free text guide.

Three things the word means

Sense	What is paired	Who uses it this way (read 11 September 2026)	Where it lives on this site
1. Flows plus a model	Scripted flows, menus and rules for known paths; an NLU classifier or an LLM for typed messages outside them	IBM ("combines rule-based logic with machine learning capabilities"); Tidio's types article; Rasa's CALM page, whose comparison table labels CALM the "Hybrid Approach"; Google, whose Dialogflow CX docs call a flow-based agent with data stores a "hybrid agent"; Landbot ("the intelligence of AI agents with the control of rule-based chatbots")	This entry; what is a chatbot §4
2. Bot plus person	An automated first line and a human agent who takes over on request or on failure	Tidio ("Hybrids also pair the bot with live chat"); IBM's rule-based section, where the bot "transfers the user to a live support agent"; most helpdesk-first vendors, who say "bot and live chat" rather than hybrid	live chat, human handoff
3. Voice plus chat	One assistant, one set of flows and knowledge, answering both spoken and typed channels	Tidio ("hybrid bots (text and voice)"); our own voice + chat hybrid guide; Voiceflow's positioning as we summarized it in the Voiceflow review	The voice + chat guide

The three are not exclusive. A Tidio deployment with Flows, Lyro and a staffed inbox is senses 1 and 2 at once, and a Voiceflow agent on a phone line and a web widget with a person at the end is all three. When a vendor page says "hybrid" without a noun after it, sense 1 is the safe reading in 2026; ask anyway.

Sense one: two layers, and a rule for which answers first

A flows-plus-model bot is not two bots. It is one conversation with a dispatcher in front of it, and the dispatcher's rule is the design decision that every vendor documents somewhere and few put on the pricing page. "First" in what follows means whose rules get first claim on choosing the reply, not which component reads the text; on several platforms a model reads every message and a flow still decides what is said. Six vendors, read on 11 September 2026:

Tidio states the order outright. Its help article on how the three modes interact says that "as a general rule, we can say that the order in which a visitor will interact is as follows: 1. Flows 2. Lyro 3. Live agents," and then refines it: proactive flows fire before a message is sent, so they win the opening turn; "when you enable Lyro, any Visitor says flows you currently have will be disabled," to stop keyword flows and Lyro's own keyword-triggered tasks colliding; if you re-enable such a flow and it shares a phrase with a Lyro task, "the flow takes priority and triggers first"; and the AI agent itself "will take over a conversation only if there is no Lyro task and no flow available to handle the conversation." A flow can even close the seam: the Disable Text Input action means "your visitors won't be able to type anything in the chat while a flow is running," so the model never gets a turn until the flow ends. That is a hybrid with the scripted layer in charge.

Google Dialogflow CX documents the same thing as an evaluation order. Its "Data store hybrid agents" page describes upgrading a flow-based agent to a "hybrid agent, which combines the power of precise conversation controls (flows, parameters, intents, conditions, transitions, and so on) with data store handler generative features," and lists how input is evaluated: "1. Parameter input while form filling. 2. Intent matches for routes in scope. 3. Data store handler with FAQ data store content. 4. Data store handler with unstructured data store content." Form first, intents second, the generative layer last. The companion page adds a detail that matters for the buttons vs free text decision: Dialogflow CX agents "always use language models for understanding end-user intention, but you can decide whether and how language models are used for agent responses." The model reads every message; the flow decides what is said back.

Rasa runs the model before the flows, and says so. In CALM, "an LLM to interpret user input and suggest the next steps" runs first on every message, producing commands; then the dialogue manager executes "structured 'Flows,' which are predefined business processes broken down into clear steps." The page's comparison table sets the "Hybrid Approach (CALM Assistants)" between "LLM-Centric" ReAct agents and "Classic" NLU bots, and its key claim is separation: "LLMs keep the conversation fluent but don't guess your business logic." Model for understanding, flows for execution, and the model picks which flow runs, including mid-flow, which is what makes it model-first by this entry's definition.

Botpress ships the seam as a node. "By default, every bot has an Autonomous Node installed in its main Workflow," and "unlike a standard Node, which executes its Cards one by one, the Autonomous Node uses a Large Language Model (LLM) to decide when to execute tools," including transitions to sub-Workflows. The docs also show the reverse setting: with "Loops until exit condition" disabled, the node "processes the user's message," acts, and "transitions to the next Node in the Workflow without responding to the user," which the page describes as "useful if you want AI to handle logic behind the scenes, but still hard-code the bot's response." The example instructions the page prints include a line worth stealing for any hybrid: "Use buttons for quick replies and easy navigation whenever possible."

Manychat and Landbot put the model inside the flow as a block. Manychat's AI Step "uses artificial intelligence to manage complex user interactions" from "a script or a set of instructions," sits in an automation like any other node ("if the AI Step is not the first step of the automation, its channel will be set automatically based on the previous node's channel"), and can write the reply into a field. Landbot's February 2026 post describes the same shape as a "structured entry layer" (welcome, name, role, "Company size (using button-based selection)") followed by an AI Agent block that "interprets intent, asks follow-up questions based on context" and "can include interactive components (such as buttons) to balance open text with structured inputs." Flow first, model in the middle, flow again for the CRM write.

Six vendors, three orders: flow-first (Tidio, Dialogflow CX), model-first (Rasa, Botpress by default), and model-as-a-step (Manychat, Landbot). None is wrong. Flow-first is the safe default when the flows carry the money (checkout, booking, lead qualification); model-first suits a support bot where most messages are questions and the flows are the exceptions. What you cannot do is leave it to the vendor's default without knowing which one you got.

Sense two: bot plus person

The second meaning needs no model layer at all: a scripted bot that can hand the conversation to a human. IBM's description of the failure mode when it cannot reads true: a rule-based bot that cannot place a request "often" transfers the user to a live agent, "but if that transfer isn't enabled, the chatbot ends up acting as a gatekeeper." Tidio's article makes the same pairing part of the definition: "the bot manages the conversation until a request gets complex, then a support agent steps in without breaking the flow."

On this site that mechanism has its own entries. The live chat entry describes the bot-first, person-second pattern most SMB deployments run; human handoff covers the transfer itself; the escalation playbook lists the triggers. What sense 2 adds to sense 1 is a fourth layer in the priority order, and Tidio's list puts it exactly there: flows, tasks, AI agent, then "Agents," who "can monitor and join conversations at will, including Lyro chats."

Sense three: voice plus chat

The third meaning is a channel pairing, not a logic pairing. Tidio's line is that voice bots "usually split into hybrid bots (text and voice) and voice-only bots"; our voice + chat hybrid guide treats it as "a single assistant that meets customers on both spoken and typed channels, ideally sharing the same flows, knowledge, and handoff logic underneath." The seam questions are different here: whether the same flow can render a menu as buttons on chat and as a spoken list on voice, and whether the no-input reprompt that voice needs is switched off on messaging channels, where a customer who has not answered in ten seconds has simply put the phone down. If a vendor says hybrid and then talks about telephony, this is the sense they mean.

Three questions that tell you which hybrid you are being sold

The word on the pricing page does not settle it. These do, and each has a documented answer on the platforms above.

1. When a message arrives, whose rules pick the reply first? Ask for the order in writing. Tidio's answer is a numbered list; Dialogflow CX's is a four-step evaluation; Rasa's is the language model first, then flows. If the answer is "the AI handles it," ask what happens to a message that arrives while a form is half filled. On Dialogflow CX the parameter wins; on Tidio the flow wins unless the visitor's typed message interrupts it; on a model-first platform the model may reinterpret the answer as a new request, which is the failure our multi-turn form guide is about.

2. Can a flow lock the model out, and can the model leave the flow? Tidio's Disable Text Input and Voiceflow's per-step setting that, when off, means "the agent only responds to button clicks" (read in the Buttons step documentation on 11 September 2026) are the first; Botpress's sub-Workflow transitions and Landbot's AI Agent block handing back to the next question are the second. A hybrid where neither is possible is two products in one login, and the customer will find the gap between them.

3. What crosses the seam? When the model layer finishes, does the flow get a usable value or a paragraph? Manychat's AI Step can "save the user's reply" to a system or custom field; Botpress gives an Autonomous Node read access to the variables you add to it and, if you check the box, write access; Landbot's post is blunt about the alternative, where "visitors type freely, the system responds in lengthy paragraphs, and important details end up buried in unstructured text." If the answer is "the transcript," you have a chat window with two brains, not a hybrid.

Why hybrids became the production default

The pairing exists because each layer alone fails in a way the other fixes. IBM's menu-based section has the scripted failure in one clause: "if a user's need isn't listed as a menu option, the chatbot will be useless because it lacks a free text input field." The model layer's failure is cost and control. Google's agent-design guidance for Dialogflow CX, read 11 September 2026, says "you should have at least 20 training phrases for every intent," and "roughly 50 is desirable" for head intents, which is the price of understanding typed text with a classifier; an LLM removes the training phrases and adds latency, a per-message bill and the hallucination problem, which is why Rasa's page sells "deterministic execution" as a feature. The generative vs rule-based guide makes the flow-by-flow call; the short version is that a hybrid keeps the money paths enumerable and testable and spends the model on the long tail, which is the same trade the buttons vs free text guide makes one question at a time.

What our fifteen reviews record about hybrids

The word appears in 4 of the fifteen review files matched by sample-reviews/*-review.md (searched 11 September 2026, case-insensitively), and only one of the four uses it in a vendor's own sense-1 positioning. The Landbot review records the homepage line "AI Agent & Chatbot Builder" with "hybrid AI + rule-based positioning" (captured 26 May 2026) and the AI product page's "the intelligence of AI with the control of rule-based chatbots," which on 11 September 2026 is still the wording of the product-menu blurb across landbot.io, while the hero subheading under that page's H1 now reads "the intelligence of AI agents with the control of rule-based chatbots." The Voiceflow review uses the word in two senses, on three lines, in our own voice: for the "voice + chat hybrid" tagline (sense 3) and, twice, for the "deterministic flow + LLM-orchestration hybrid" it found in the agent model (sense 1). The AiSensy review uses it for "hybrid retrieval" inside a knowledge base, a search-engineering sense that is not a chatbot architecture, and the Blip review uses it for a version label of our own methodology. So one vendor claim, two editorial descriptions, two unrelated uses.

What the reviews do record, without the word, is the seam. The Tidio review built a "5-step Quick-Replies Flow" in nine minutes and separately measured Lyro on a fifteen-question test against a single-page support index (both on 28 May 2026); the Manychat review is the one file that names the AI Step, and records it as available on the Free tier; the Landbot review captured the AI Agent block with its Role, Constraints, Context and Steps instruction scaffold; the AiSensy review found Knowledge Base, AI Orchestrator and Tool Calling as tabs inside the flow builder. None of the fifteen protocols asked the vendor for the priority order in writing or tested a typed message arriving mid-form; that is a gap in our protocol, recorded here, and the three questions above are what a future round should ask. Corrections to editorial@chatbotscape.com.

What is a chatbot — the four-architecture taxonomy in which the hybrid is the fourth pattern.
Generative AI chatbot — the model layer on its own, and what it costs to run.
Fallback intent — where a typed message goes when the flow layer cannot place it and there is no model layer.
Human handoff — the transfer that sense 2 is named for.
Live chat — the bot-first, person-second pattern most SMB deployments run.
Intent recognition — the classifier that was the model layer before LLMs, and still is on several platforms.

FAQ

What is a hybrid chatbot?

A bot that runs a scripted flow layer (buttons, menus, keyword rules, forms) and a model layer (an NLU classifier or a large language model) under one conversation, with a rule for which answers first. IBM's definition is "a conversational AI system that combines rule-based logic with machine learning capabilities." Vendors also use the word for a bot paired with human agents and for one bot serving voice and chat, so ask which pairing is meant.

Is a hybrid chatbot the same as an AI chatbot?

No. A fully generative AI chatbot answers everything with the model; a hybrid keeps the predictable paths (order status, booking, lead capture) as flows the model cannot rewrite and sends only the unanticipated messages to the model. Dialogflow CX's docs draw the line precisely: language models always interpret what the user meant, but "you can decide whether and how language models are used for agent responses."

Which chatbot platforms are hybrid?

In the flows-plus-model sense, most of the SMB builders we review: Tidio (Flows plus Lyro, flows first), Manychat (AI Step inside an automation), Landbot (AI Agent block inside a flow), Botpress (an Autonomous Node in the main workflow with sub-Workflows for fixed paths), Voiceflow, and on the developer side Rasa (CALM) and Google Dialogflow CX (flows with data stores or generative fallback). The order in which the layers run differs, and that is the thing to ask about.

What is the difference between a hybrid chatbot and a rule-based chatbot?

A rule-based bot has only the flow layer: if the customer's message matches no button, keyword or menu option, it falls to a fallback intent or a person. A hybrid adds a model that reads the message instead. The generative vs rule-based guide decides, flow by flow, which paths are worth moving.

Does a hybrid chatbot still need buttons?

Yes, and the vendors that ship the model layer say so themselves: Botpress's example instructions tell the Autonomous Node to "use buttons for quick replies and easy navigation whenever possible," and Landbot's AI Agent "introduces selectable options or buttons to simplify interaction and reduce cognitive load" when appropriate. Buttons make the flow layer's answers enumerable and cheap; the model is for the questions no button anticipated. The per-question rule is in the buttons vs free text guide.

Sources

IBM, 6 types of chatbots and how to choose the right one for your business (ibm.com/think/topics/chatbot-types, by Bella Church and Teaganne Finn, dated 7 March 2025), read 11 September 2026: the hybrid definition ("A hybrid chatbot is a conversational AI system that combines rule-based logic with machine learning capabilities"; "The hybrid chatbot provides the best of both systems in one single system"); the menu-based section ("Essentially, these chatbots operate like a decision tree"; "if a user's need isn't listed as a menu option, the chatbot will be useless because it lacks a free text input field"); the rule-based section's transfer sentence ("Often, the chatbot transfers the user to a live support agent, but if that transfer isn't enabled, the chatbot ends up acting as a gatekeeper").
Tidio, Types of Chatbots: 7 Main Categories Explained (2026) (tidio.com/blog/chatbot-types, by Polina Fomenkova, updated 2 July 2026), read 11 September 2026: "Hybrid chatbots combine a rule-based layer with an AI layer, so they can run fast scripted flows and still handle open-ended questions"; "Hybrids also pair the bot with live chat. The bot manages the conversation until a request gets complex, then a support agent steps in without breaking the flow"; "Voice bots usually split into hybrid bots (text and voice) and voice-only bots." A vendor article; used here for the three senses, not for its product claims.
Tidio Help, How do flows, Lyro, and live agents work together? (help.tidio.com, article 14667264947356, dated 4 September 2025), read 11 September 2026: the general order ("1. Flows 2. Lyro 3. Live agents"); "A proactive flow will trigger before a visitor sends a chat message, which allows the flow to take priority over Lyro"; "when you enable Lyro, any Visitor says flows you currently have will be disabled"; "If a Visitor says flow shares keywords/phrases with a Lyro task, the flow takes priority and triggers first"; "Lyro will take over a conversation only if there is no Lyro task and no flow available to handle the conversation"; agents "can monitor and join conversations at will, including Lyro chats." Tidio Help, Disable Text Input (article 5465426393756, dated 15 May 2025), read 11 September 2026: "your visitors won't be able to type anything in the chat while a flow is running."
Google Cloud, Data store hybrid agents, Dialogflow CX documentation (docs.cloud.google.com/dialogflow/cx/docs/concept/data-store/hybrid; stamped "Last updated 2026-07-29 UTC" in the copy fetched first and "Last updated 2026-09-03 UTC" when re-read in a second session about an hour later on the same day, with the quoted text identical in both), read 11 September 2026: the hybrid-agent definition and the four-step "Input evaluation order" quoted above. Google Cloud, Generative versus deterministic (docs.cloud.google.com/dialogflow/cx/docs/generative-deterministic, "Last updated 2026-09-03 UTC"), read 11 September 2026: "Dialogflow CX agents always use language models for understanding end-user intention, but you can decide whether and how language models are used for agent responses"; the fully generative, partly generative and deterministic feature groups. Google Cloud, General agent design best practices (docs.cloud.google.com/dialogflow/cx/docs/concept/agent-design, "Last updated 2026-09-03 UTC"), read 11 September 2026: "You should have at least 20 training phrases for every intent"; "roughly 50 is desirable" for head intents of large agents.
Rasa Documentation, Conversational AI with Language Models (CALM Overview) (rasa.com/docs/learn/concepts/calm, "Last updated on Sep 10, 2026"), read 11 September 2026: "By combining language model flexibility with predefined logic"; "LLMs keep the conversation fluent but don't guess your business logic"; "CALM is a controlled framework that uses an LLM to interpret user input and suggest the next steps"; "structured 'Flows,' which are predefined business processes broken down into clear steps"; the comparison table's three columns, "LLM-Centric Approach (ReAct-Style Agents)," "Hybrid Approach (CALM Assistants)" and "Classic Approach (NLU Bots)"; "Deterministic execution" among the key benefits.
Botpress Documentation, Autonomous Node (botpress.com/docs/studio/concepts/nodes/autonomous-node, no date on the page), read 11 September 2026: "By default, every bot has an Autonomous Node installed in its main Workflow"; "Unlike a standard Node, which executes its Cards one by one, the Autonomous Node uses a Large Language Model (LLM) to decide when to execute tools"; the non-conversational setting ("transitions to the next Node in the Workflow without responding to the user"; "useful if you want AI to handle logic behind the scenes, but still hard-code the bot's response"); variable access ("By default, Autonomous Nodes can't read or update variables"; "Allow write access"); the example instructions line "Use buttons for quick replies and easy navigation whenever possible."
Manychat Help, Manychat AI Step (help.manychat.com, article 14281187288860, updated 31 July 2026), read 11 September 2026: "uses artificial intelligence to manage complex user interactions. It operates based on a script or a set of instructions"; "If the AI Step is not the first step of the automation, its channel will be set automatically based on the previous node's channel"; the option "to save the user's reply" to a System or Custom Field; channel support (Messenger, Instagram, WhatsApp, TikTok, Telegram).
Landbot, AI Agent Chatbots (landbot.io/ai-agent-chatbots, no date on the page), read 11 September 2026: the hero subheading beneath the H1, "Combine the intelligence of AI agents with the control of rule-based chatbots"; the site-wide product-menu blurb "Combine the intelligence of AI with the control of rule-based chatbots to build a enjoyable conversational experiences" (sic); "Blend AI with static flows." Landbot, Is Landbot a Chatbot or an AI Agent? (landbot.io/blog/is-landbot-a-chatbot-or-an-ai-agent, by Irene Pomares, 20 February 2026), read 11 September 2026: the "structured entry layer" with "Company size (using button-based selection)"; the AI Agent layer that "interprets intent, asks follow-up questions based on context" and "can include interactive components (such as buttons) to balance open text with structured inputs"; "Visitors type freely, the system responds in lengthy paragraphs, and important details end up buried in unstructured text"; "When appropriate, the AI Agent introduces selectable options or buttons to simplify interaction and reduce cognitive load."
Voiceflow Documentation, Buttons (docs.voiceflow.com/documentation/buttons/steps/buttons, which redirects to the same page under voiceflow.com/docs/documentation/build/steps/buttons; no date in the Markdown rendering), read 11 September 2026: the "Listen for other triggers" setting, "When disabled, the agent only responds to button clicks."
Chatbotscape review corpus (the fifteen platform reviews listed at /reviews), searched 11 September 2026 from the repository root. Denominator: ls sample-reviews/*-review.md | wc -l returns 15. grep -liE 'hybrid' sample-reviews/*-review.md returns 4 (aisensy, blip, landbot, voiceflow); grep -liE 'AI Step' sample-reviews/*-review.md returns 1 (manychat); grep -liE 'Autonomous Node' sample-reviews/*-review.md returns 0, which is why Botpress's seam is cited to its documentation above and not to our review; grep -liE 'flow builder' sample-reviews/*-review.md | wc -l returns 13 (the two files without the phrase are chatbase and tars; Tars sells a flow builder its review never calls by that name, so this is a string match, not a product count) and grep -liE 'knowledge base' sample-reviews/*-review.md | wc -l returns 15. Passages cited: landbot-review.md lines 124, 152, 230 and 689–702 (homepage and AI-page positioning captured 26 May 2026; the AI Agent instruction scaffold); voiceflow-review.md lines 38, 208 and 357; aisensy-review.md lines 233 and 714; blip-review.md line 306; tidio-review.md lines 205 and 223; manychat-review.md lines 144 and 619–631 (the AI Step in Flow Builder and its Free-tier availability).
Ahrefs Keywords Explorer, US overview, queried 10 September 2026 — the demand, difficulty, CPC, global-volume and parent-topic figures in this entry's keyword note.
Chatbotscape evaluation methodology. /methodology (continuously updated).