Skip to main content

​
A

Adapter A provider implemented first-party rather than through Pipecat, living under apps/providers/adapters/. Two exist today: Bhashini (TTS) and Kenpath (LLM). Reported as provider_type: adapter. See Provider registry. Agent A configured voice assistant: prompts, behaviour, language, and its STT, TTS, and LLM choices. Belongs to one organisation. See Agents. Agent category Either telephony (reached over the phone, provisioned with a provider application) or websocket (reached from a browser). Determines whether /answer works and which frame serializer the runtime uses. Application A configuration object created on a telephony provider’s account that ties a phone number to VoicEra’s answer URL. The API provisions and deletes these for you. ARQ The Redis-backed job queue running campaign batches and CSV syncs off the request path. See Workers and orchestrator.

​
B

Barge-in A caller interrupting the agent mid-sentence. The pipeline stops playback and starts listening. Tunable with interruption_min_words. See Voice pipeline. Bot token A short-lived, organisation-scoped JWT minted by POST /users/bot/token against the INTERNAL_API_KEY. How the runtime authenticates to the API. See Multi-tenancy.

​
C

Call log The record of one call: type, status, response, timings, and pointers to its transcript and recording. Stored in CallLogs. Campaign A CSV-driven outbound calling run with retry policy, scheduling, concurrency limits, and a circuit breaker. See Campaigns. Circuit breaker A guard that pauses a campaign when the failure rate in a rolling window crosses a threshold, so a broken configuration cannot burn an entire contact list. Chroma The embedded vector store holding per-organisation RAG chunks, persisted to the voicera_oss_chroma_data volume. Concurrency slot A Redis-held token representing one in-flight call. Bounded per organisation and per campaign. See Call concurrency.

​
D

Discriminated union The Pydantic pattern in apps/providers/factory.py that turns the registry into a single validated type, so adding a provider never means editing an if/elif chain.

​
F

FerretDB A proxy speaking the MongoDB wire protocol on top of PostgreSQL. VoicEra’s data store. Published on host port 27018. See Data store. Frame Pipecat’s unit of data flowing through the pipeline: an audio chunk, a transcript, an LLM token, a control signal. Frame serializer The translator between a telephony provider’s wire format and Pipecat frames. One per provider, registered in apps/telephony.

​
G

Gateway The single published entry point of the model server, on port 8100. Routes to whichever STT, TTS, and LLM slots are deployed. See Gateway API.

​
H

Hold message Audio played when the agent needs time — typically while a tool call or retrieval runs. Configured in the agent’s behaviour block.

​
K

Knowledge base Documents ingested, chunked, embedded, and retrieved to ground an agent’s answers. Two modes, tool and context. See Knowledge base.

​
M

Membership The join record between a user and an organisation, carrying the role. See Multi-tenancy. MinIO The S3-compatible object store holding call recordings and transcripts. Model server The optional self-hosted stack running STT, TTS, and LLM on your own hardware behind one gateway. See Model server.

​
O

Orchestrator The campaign-orchestrator container. Listens on Redis pub/sub, schedules the next campaign batch, and detects completion. Organisation The tenant boundary. Every agent, number, credential, campaign, and call belongs to exactly one.

​
P

Partial transcript Words returned while the speaker is still talking. Required for responsive telephony. Distinct from whether the model serves a streaming endpoint — see STT models. Pipecat The async Python framework for real-time voice pipelines that the runtime is built on. ProviderAuth The Fernet-encrypted, provider-level credential record for an organisation. Replaces the old Integrations model. See Provider credentials. Provider registry The self-describing catalogue of STT, TTS, and LLM vendors. Providers register themselves; the schema dump drives client forms. See Provider registry.

​
Q

QueuedRun One queued outbound call in a campaign — a single contact and retry attempt. Stored in QueuedRuns.

​
R

RAG Retrieval-augmented generation: fetching relevant chunks from your documents and giving them to the LLM as grounding. Runtime The apps/runtime service on port 7860. Serves the /answer webhook and runs one Pipecat pipeline per call. See Runtime. RTVI The real-time voice interaction protocol used by Pipecat browser clients over protobuf frames.

​
S

Slot One of the model server’s three fixed positions — stt, tts, llm — each a container on a fixed internal port. Which model fills a slot is a folder name in .env. See Slots and models. Stream XML The XML returned by /answer telling a telephony provider which WebSocket URL to stream the call audio to.

​
T

Telephony provider Vobiz or Plivo. Owns the phone numbers and streams call audio. Selected per agent, never global. See Telephony model.

​
V

VAD Voice activity detection — deciding when the caller starts and stops speaking, which drives turn-taking and barge-in.