Terms are grouped by area. API field names are in code font.
Agents and engines#
| Term | Meaning |
|---|---|
| Agent | The resource a call runs on. Holds voice, language, speech, call, analysis, privacy and webhook settings, and points at a response engine. ID prefix agent_. |
| Response engine | What produces the agent's replies: a Retell LLM (retell-llm), a conversation flow (conversation-flow) or your server (custom-llm). Set in response_engine. |
| Retell LLM | The prompt engine for single and multi-prompt agents: model, prompt, tools, states, knowledge bases. ID prefix llm_. The name matches Retell's API. |
| Single prompt agent | A Retell LLM with one general_prompt and no states. |
| Multi-prompt agent | A Retell LLM with states. Each state has its own prompt, tools and edges. |
| State | One stage of a multi-prompt agent. The model moves between states with transition_to_<state> functions. |
| Conversation flow | A graph of nodes and edges that drives the call. ID prefix conversation_flow_. |
| Node | One step in a flow: conversation, function, transfer, press digit, branch, extract variables, SMS, code, MCP, agent swap, component, end. |
| Edge | A transition between nodes with a prompt, equation or combined condition. |
| Global node | A flow node that can be entered from any point when its condition holds. |
| Component | A reusable subflow, stored in the flow or in the shared library. |
| Flex mode | A flow compiled into one prompt so the model moves between steps itself (flex_mode). |
| Custom LLM | An agent whose replies come from your WebSocket server at {llm_websocket_url}/{call_id}. |
| Chat agent | A text agent for the chat API, website widget and SMS. Uses a Retell LLM or flow engine. |
| Transfer agent | An agent used for agentic warm transfers: it briefs the destination and then bridges or cancels the transfer (is_transfer_agent). |
| Template | A starting prompt, voice, functions and post-call fields in the Create agent dialog. |
Conversation#
| Term | Meaning |
|---|---|
| Turn | One speaker's contribution: the caller speaks, then the agent replies. |
| Turn-taking | Deciding when the caller has finished so the agent can speak. |
| Endpointing | The silence after speech before a turn or transcript is treated as finished. |
| Turn detector | A model that judges whether an utterance is complete, so the agent need not wait the full silence window. |
| Responsiveness | How quickly the agent replies after the caller stops (responsiveness, 0 to 1). |
| Interruption sensitivity | How easily the caller can talk over the agent (interruption_sensitivity, 0 to 1; 0 means no interruptions). |
| Preemptive generation | Starting the LLM on the interim transcript before the turn is confirmed. |
| Backchannel | Short words like "mm-hmm" the agent says while the caller talks (enable_backchannel). |
| Reminder | A nudge when the caller goes quiet (reminder_trigger_ms, reminder_max_count). |
| Welcome message | The agent's first line. Static text, generated from the prompt, or nothing when the user speaks first (begin_message, start_speaker). |
| Function, tool | An action the model can call: end call, transfer, press digit, custom HTTP, code, extract variables, SMS, MCP, integration app, agent transfer. |
| Dynamic variable | A {{name}} placeholder filled with a string at run time. |
| Collected variables | Values the agent captured during the call, stored in collected_dynamic_variables. |
| System variable | A variable Tried fills itself, such as {{current_time}} or {{user_number}}. |
| Knowledge base | Documents, web pages and text that are chunked, embedded and searched every turn. ID prefix knowledge_base_. |
| Chunk | A piece of a knowledge base source returned by search. |
| Agent Handbook | Preset behaviours added to the prompt, such as echo verification or AI disclosure (handbook_config). |
| Guardrails | Topic filters on what the agent says and on caller jailbreak or scam attempts (guardrail_config). |
| Contact memory | A rolling summary of past conversations with a contact, read into calls as {{contact_memory}}. |
Speech#
| Term | Meaning |
|---|---|
| STT | Speech to text: the live transcription of the caller. Deepgram by default. |
| TTS | Text to speech: the agent's voice. |
| LLM | The language model that writes the replies. gpt-5.6-terra by default. |
| Speech-to-speech | A model that takes audio in and returns audio out, replacing STT, LLM and TTS (s2s_model). |
| Voice tier | Premium (default) or Standard, stored as the agent's voice_model. See Voice. |
| Fallback voices | Voices tried in order when the primary voice provider fails (fallback_voice_ids). |
| Boosted keywords | Words the transcriber should favour, such as brand names (boosted_keywords). |
| Denoising | Removing background noise, and optionally background voices, from the caller audio (denoising_mode). |
| Pronunciation dictionary | Word to phoneme rules for the voice (pronunciation_dictionary). |
| Expressive mode | Lets supported voices perform tags such as [laughs] (enable_expressive_mode). |
Calls and telephony#
| Term | Meaning |
|---|---|
| Call | One phone or web conversation. ID prefix call_. Status registered, ongoing, ended, error or not_connected. |
| Phone call | A call over the phone network, call_type: "phone_call". Direction inbound or outbound. |
| Web call | A call from a browser over WebRTC, call_type: "web_call". |
| Disconnection reason | Why the call ended, such as user_hangup, agent_hangup, voicemail_reached or inactivity. |
| SIP trunk | A connection to a carrier or PBX that carries calls to and from Tried. |
| Inbound webhook | A URL called when a number receives a call, before the agent answers. It can change the agent or variables. |
| Cold transfer | Hand the caller to a number without introducing them. |
| Warm transfer | The agent briefs the person at the destination before connecting the caller. |
| DTMF | Keypad tones. Callers can enter digits, and the agent can press digits in phone menus. |
| IVR | An automated phone menu ("press 1 for sales"). |
| Voicemail detection | Detecting an answering machine on outbound calls and hanging up or leaving a message. |
| Batch call | Many outbound calls from one list, with a schedule and time window. ID prefix batch_call_. |
| Concurrency | How many calls a workspace can run at once. |
| A/B testing | Splitting a number's calls across agents by weight. |
Versions and analysis#
| Term | Meaning |
|---|---|
| Draft | An editable version of an agent. |
| Published version | A frozen version. Edits create a new draft. |
| Environment tag | A named pointer such as prod or staging to a published version, with its own dynamic variable values. |
| Post-call analysis | An LLM pass after the call that fills call_summary, call_successful, user_sentiment, in_voicemail and your custom fields. |
| Post-Call Extraction | The dashboard section where you define custom analysis fields. |
| Workflow | Steps that run before dispatch (pre-call) or after analysis (post-call). |
| Webhook | An HTTP POST Tried sends to your server on call events, signed with x-tried-signature. |
| Conductor | The dashboard's AI assistant for building agents, writing tests and investigating calls. |
| Mock mode | Each provider runs a built-in mock until its key is set, so the platform works without keys. |