Skip to main content

Prompt, models and knowledge

Agent & prompt

  • Global prompt · personality & rules — the agent’s personality, rules, tone and general behaviour. It applies across the whole conversation. The button to the right of the label opens the prompt in a full-screen editor: it shows the agent’s language and model, a character count and the help from «?». Save or ⌘ Enter writes the text back to the field, Cancel or Esc leaves the field as it was.
  • Agent languages — the languages the agent speaks in; the cabinet language does not change with them. See Agent languages below.
  • Execution model — see Execution model below.
  • History window — how many recent messages are sent to the model — is no longer set on the agent: it is a platform setting, see Platform settings for agents.
  • Entry point — set by the Begin node on the canvas: its single edge leads to the first node of the dialogue, and who speaks first and the start after caller silence are set there too. Details — Begin node.
The Agent & prompt card: global prompt, agent languages and execution model

Agent languages

The Agent languages field is the list of languages the agent talks in. The first language is marked Primary: every call starts in it.
  • A click on the field — on a chip, an empty spot or the arrow on the right — opens the list of the platform’s language catalog with a search; another click closes it. Tick the languages you need; Make primary moves a language to the top of the list, Clear unticks everything, Done closes the list.
  • The cross on a language removes it from the list. The last language cannot be removed: without a language the agent cannot speak.
  • Auto-switch to the caller’s language — the switch appears when there are two languages or more; on by default. On — the agent detects which of its languages the caller speaks and answers in it. Off — the call stays in the primary language until a scenario step switches it (the Call language group in the Override Settings node).
The call language. Everything the agent says and hears follows the current language of the call:
  • the voice speaks the current language, and recognition listens to all agent languages — nothing to switch separately;
  • the built-in phrases — the silence reminder, the wait acknowledgement, the active-listening words, the transfer intro, the apology on a model failure — are said in the current call language;
  • if the caller speaks a language that is not in the agent’s list, the call stays in the primary language;
  • a switch shows in the call history as a Call language event.
The voice, the recognition model and the realtime model must support every agent language. The pickers show only the options that fit. If the chosen model or voice stopped fitting after a language was added, it stays first in the list marked Does not support: …, and publishing is blocked until you pick another one or remove the language.
“Multiple languages” is no longer an agent language. If the agent has this option, the field asks you to pick the languages as a list: without that the agent cannot be published.
Every language of an agent imported from a file or sent through the API must be a code of the language catalog, letter for letter: kz, KK or kk-KZ instead of kk blocks publishing. The error names the right code; pick the language again in the Agent languages field.

Variables in authored text

A {{variable_name}} placeholder is replaced with the session value everywhere you write text, not only in a node instruction:
  • the agent’s Global prompt;
  • transition conditions, including global ones;
  • the agent objectives summary;
  • tool and function descriptions and arguments;
  • variable extraction fields;
  • the flex-mode prompt;
  • the answering-machine detection hint;
  • fine-tuning example texts.
Substitution covers the global prompt, not only the node instruction. A script placed in the global prompt reaches the model with its values filled in — there is no need to copy its text into node instructions to get substitution.
An unknown name stays as text rather than becoming emptiness. If the variable is not present on the call, the model sees {{name}} — exactly as this has always worked in a node instruction. An empty substitution would make the author’s typo invisible: the agent would quietly speak with a hole in the sentence instead of a clear signal that the variable never arrived.Missing file columns are caught elsewhere — at the campaign launch step, and only for the variables ticked In signature on the agent (see Agent variables): a signature variable with no column blocks the launch.
Agents without placeholders are not affected at all. Text with no {{ renders to itself byte for byte, so the prompt prefix cache and its warm-up work at full strength for such agents. For an agent that does use substitution the prefix differs on every call — and warm-up skips it: warming text that will never be requested would only inflate the hit statistics.
When the model refuses to answer, the agent apologizes in the language of the call. If the language model declines a turn, the agent says a short fallback line instead of going silent. The line is said in the current call language; if there is no built-in line for it — in the agent’s primary language, and if there is none for that either — in English. Choosing it costs the caller nothing — no extra turn and no network call.

Execution model

The Execution model field in the Agent & prompt card controls the model used for the agent’s main runtime operations: reply generation, tool calls and variable extraction — for every node of the agent. The priority is node → agent → server: a model set on a node wins over the agent’s, and the agent’s wins over the server configuration.
  • Default in the field means the agent has no model of its own and uses the server configuration.
  • Picking a model from the model directory sets it for this agent — there is no separate override switch.
  • Reset to default next to the field drops the chosen model and returns the agent to the server configuration. The gear parameters stay.
  • The gear opens Execution model — parameters: Temperature and Answer limit (the maximum number of tokens in a reply; empty — the platform value), Service tier (see below) and, for a reasoning model, Reasoning effort (low / medium / high). The field bounds come from the description of the selected model; a value outside them blocks publishing the agent.
In realtime mode the field is not shown: the conversation is run by the realtime model. Models come from the catalog only. The address and the key of the execution, classifier, post-call analytics, observer, recognition and voice models come from the catalog record; an agent cannot set its own address or key. If an agent still has a model set by hand, the field shows Outside the directory, and the check does not let you publish the agent or start a test: “slot … is set by hand (…): the address and the key come from the catalog only — pick a record”. Pick a catalog record, and the agent publishes again.

Agent classifier model

Request understanding sets the classifier model used for per-turn transition or stay decisions. The classifier decides whether the conversation should move to another node or remain on the current one. The section lives in Additional → Models & routing. The field works like the execution model: Default means the server classifier is used, picking a model from the directory sets it for this agent, Reset to default returns to the server one, and the gear holds the parameters. This is a latency-sensitive part of the pipeline, so a fast non-reasoning model is usually a better choice. The agent’s classifier model can be replaced on the canvas: a Logic Split node has its own Transition decision model for choosing its edge, and the classifier group of an Override Settings node sets the model for nodes without one of their own until the call ends. See Logic Split node and Override Settings node.

Reserve services

If a primary service does not answer, the agent switches to the reserve and keeps the conversation going. The reserves of all four services are set in one table: Additional → Fallback → Reserve services. The rows are Answer model (the execution model), Request understanding (the classifier model), Speech recognition and Voice. The Primary column shows the agent’s service, the Reserve column holds the reserve picker. The Reserve services section: four rows with Primary and Reserve columns and reserves from other providers
  • A reserve is always another provider. Models and voices of the primary service’s provider are not offered as a reserve: when a provider is down, all its models and voices are down. An empty field is Not chosen; the last item of the list is No reserve.
  • Above the table is a warning “No reserve: … If they fail, the call is dropped.” listing the rows without a reserve. Under the voice there is a warning that the voice will change in the middle of the call and the caller will notice.
  • In the Agent & prompt card, under the execution model, the line Reserve: … · change or No reserve set · set opens this section.
  • The agent check warns when the reserve voice is from the same provider as the primary one. The button in the warning opens the section.
  • In realtime mode reserves do not apply, and the section says so instead of the table.
How each reserve works:
  • Answer model — the reserve serves the turn on which the primary model did not start answering in time, and every new turn while the primary model is unavailable. If a node has its own model of the same provider as the agent’s reserve, the reserve does not work on that node — the agent check warns about it.
  • Request understanding — the same for the classifier model.
  • Speech recognition — the call moves to the reserve at the first recognition failure. While the primary provider is unavailable, new calls start on the reserve.
  • Voice — the phrase the primary voice failed on is spoken by the reserve voice, and the call goes on with it without returning to the primary. If a synthesis provider cannot say a single phrase for 2 minutes, the platform treats it as unavailable: new calls start on the reserve right away until the provider answers again without failures for 3 minutes in a row. Calls in progress do not switch. Cybernet voices switch to the reserve only within a call. The reserve voice sounds the way the directory sets it — without the agent’s model, speed and fine-tuning.
The switch to a reserve shows in the call card.

Decision model for transitions (Jev)

The transition decision can be taken by a decision model — a System One model (Jev) — instead of the LLM classifier. It does not generate text: it picks one of the transition options and reports how confident it is. The setting lives in Additional → Models & routing → Transition classifier. The Mode is chosen from four options:
  • As on the server — the default. The agent follows the platform default; by default the platform has the mode off.
  • Off — the LLM classifier decides, as before.
  • Shadow — every decision is also sent to the decision model in the background. Its answer and confidence go into the Transition decision call event, while the route and the turn time stay the same. Use it to compare decisions on your own calls without changing anything for the caller.
  • Cascade — the decision model decides when its confidence reaches the threshold. Below it, on timeout or on error the LLM classifier decides with the same input as in the off mode.
The full description of each mode is behind «?» next to Mode. Unless the mode is off, more fields appear:
  • Decision model — a record of the model directory with the role “decision”. In shadow and cascade modes publishing is blocked until a record is selected. The record is frozen into the published version.
  • Confidence threshold — cascade only: at this confidence or above the decision model decides. In shadow mode the field is locked — the server threshold applies. Left empty — the platform value (0.8 by default).
  • Timeout, ms — in cascade, how long a turn waits for the decision model before the LLM classifier takes over; in shadow mode, how long the background call may run. Left empty — the platform value (350 ms by default).
Turn cascade on only after shadow mode. First check on live calls how often the model’s decisions agree with the LLM classifier, where they differ, and what the latency is. In FLEX mode the decision model is not used: the transition there is part of the reply.

Service tier

A model slot — both the agent’s and one set on a node — has a Service tier select: the priority the provider serves the request with. On the agent it is in the gear of the model field. This is not a different model, it is the same model served with a different priority:
  • Provider default (not sent) — no tier is requested at all, so behaviour is unchanged;
  • priority — faster;
  • flex — slower;
  • auto — the choice is left to the provider;
  • default — an explicit request for the provider’s ordinary tier.
The select appears only on models carrying the Supports service tiers (OpenAI service_tier) flag; for the rest it is absent. A hint next to the select warns about poor combinations: flex in a live conversational turn makes the caller wait longer, while priority where nobody is waiting for the answer — background analytics — gains nothing. It is a hint, not a restriction.

Platform settings for agents

Two parameters are not set on the agent: the platform sets them, the same for all agents. They apply to published versions too, without a republish.
  • History window (messages) — how many recent messages of the dialog are sent to the model: 4 to 400, 40 by default (roughly 4–5 minutes of conversation). The classifier gets the same slice, so a large window adds latency to every turn.
  • Query rewrite — the knowledge-base query rewrite mode for every agent and node set to “Default”: LLM rewrite (the default) or off. An agent or a node can still switch it off on its own — see Search settings.
The rewrite model is set by the platform as well. If it is not set, the agent’s classifier model is used — with its endpoint and key.
Values set on the agent earlier no longer apply. If an agent had its own history window or its own rewrite model, it now works with the platform values. For example, an agent that had a window of 120 messages now sees the last 40 — unless the platform is set to a different value.

LLM warm-up

LLM warm-up keeps the agent’s system prompt prefixes warm in the model provider’s KV cache. This can reduce the amount of context that must be processed on the first turn of a call, so the agent can respond sooner. The section is in Additional → Models & routing → LLM warm-up.
  • The switch in the section header enables the periodic warm-up requests. Turn it off when the extra traffic is not wanted.
  • Roles — which prompt to keep cached: execution is the agent’s system prompt, classifier is the classifier prompt of the entry node.
  • Warm-up interval, sec — how often the requests go out. An empty field means the server default.
Warm-up only works with models from the model directory. When warm-up is turned off on the platform, the section is inactive and reads “LLM warm-up is turned off on this installation. The agent’s settings are kept and apply again once it is turned back on.” The Additional window open on the LLM warm-up section

Knowledge bases (RAG)

The agent has a default knowledge base set that is used by every Conversation node unless that node has its own knowledge base binding. The set is edited in the Knowledge bases card of the tab.
  • Adding a base — click the Search knowledge bases… field: the list shows every base not yet attached to the agent. Typing narrows it by name; Enter takes the first match. A chosen base appears under the field.
  • Removing a base — click its name with ✕ under the field.
  • Creating a base — not in the card: bases are created on the Knowledge bases page and then picked here.
  • An empty set — search then runs only on the nodes that define a set of their own. The card says so while no base is attached.
A base is filled with files. Formats: .pdf, .txt and .json only. Up to 5 MB per file. You can upload up to 32 files at once, but no more than 5 MB in total. Files in other formats uploaded earlier stay in the base.A base does not take a website address: to have the agent answer from the content of your pages, save them as PDF or copy the text into a .txt file and upload the files. An update at the source does not reach the base on its own: upload the new version of the file and delete the old one.How long processing takes is not known in advance: the time grows with the queue. The status of each file is shown in the base’s file table. Slow processing and a retry are not errors; a file is marked as failed by time only if it has not been processed within 2 hours from the start of processing. A base with only some files ready already opens and works with the ready ones.To get a file back, press Download file to the left of the delete button in the base’s file table: the original is saved under the name shown in the list, byte for byte as it was uploaded. While the file is still being processed the button is disabled with the hint “The file is still being processed”; for ready and failed files it is active. Every download is written to the audit log.
Finding a base. On the Knowledge bases page, the Search by name or ID field filters the list by part of the name or of the full ID, ignoring case. The search is kept in the page address, so a link opens the list already filtered. When nothing matches, the page says Nothing found and offers to reset the search. The ID column shows the first 8 characters; hover to see the full ID, and Copy ID copies it without opening the base. Inside a base the full ID is shown with the same button. The Knowledge bases card with an attached base and the Search settings gear

Search settings

Search settings open from the gear in the Knowledge bases card header. An empty field means the knowledge base service default (hybrid_search / 0.4 / 5). Save writes the values to the draft, Cancel drops them; a dot on the gear means a value differs from the default.
  • Search type — how the selected bases are searched; the default (hybrid) value combines several retrieval signals.
  • Similarity threshold (0..1) — the minimum relevance a result must reach. A higher value returns fewer but more precise hits.
  • Max results (1..100) — the ceiling on the number of hits.
  • Query rewrite (LLM) — whether the query is rewritten before the search: the default item, Off (gate + merge) or LLM rewrite. The default item is Default (platform): the mode comes from the platform setting. The mode is resolved along the chain node → agent → platform.

Query rewrite model

With LLM rewrite there is one more model call before every search; on a voice turn it is capped at 2 seconds. The model is set not on the agent but on the platform — the Query rewrite model slot, see Platform settings for agents. If the slot is empty, the agent’s classifier model is used.
On the platform the rewrite is on by default. Agents with a knowledge base whose search settings say “Default” make this call on every turn with a search. If response speed matters more than retrieval quality, pick Off (gate + merge) on the agent or the node.

Flex mode

Flex mode sends the whole flow to the model as one structured prompt instead of processing each part only through the standard node flow. The section is in Additional → Models & routing → Flex mode.
  • The switch in the section header — enables processing of the complete agent flow as a single structured prompt.
  • Component summary — when enabled, the section shows how many nodes are included and approximately how many prompt tokens they use, for example Component: 2 nodes, ~389 prompt tokens. Warnings about a component are shown under its line.
Flex mode section with component summary Use the summary to understand the size of the generated prompt before running the agent. Leave Flex mode disabled when the standard node-by-node flow is the better fit for the agent.