Understand data flows and stored records
The data path changes with the selected architecture, model, voice, knowledge, and functions. Review the complete configuration rather than assuming one universal provider flow.
Common stored records
| Area | Examples of stored data |
|---|---|
| Account and workspace | User ID, email and profile metadata, workspace settings, membership, role, invitations, and branding. |
| Agent configuration | Prompt, variables, model, voice, speech controls, knowledge selection, functions, drafts, and published versions. |
| Voice conversations | Caller and destination numbers, timestamps, duration, status, transcript, outcome, summary, captured caller details, and credit data. |
| Browser voice tests | Test call, test lead/contact representation, transcript, duration, and credits. |
| Beta Chat | Contact/pre-chat values, privacy-consent record when accepted, messages, tool events, tokens, errors, and credits. |
| Knowledge | Source names and URLs, extracted text chunks, embeddings, metadata, and instructions. |
| Vox | Threads, messages, research context, action records, usage, and temporary attachments. |
| Integrations | Provider, connected account metadata, configuration, and encrypted credential references. |
| Billing and operations | Subscription state, balances, usage ledger events, call attribution, and provider identifiers needed for operation. |
Do not assume that a label such as recording means an audio recording exists for every call. The current call flow reliably stores transcripts and metadata; the interface does not expose recording download, and audio-recording availability is not guaranteed.
Voice-provider flow
For a Modular voice agent:
- telephony or browser transport sends audio to the selected speech-recognition provider;
- recognized text, compiled instructions, retrieved context, conversation history, and tool definitions are sent to the selected language-model provider;
- generated response text is sent to the selected speech-synthesis provider; and
- Wirevox stores the resulting transcript and operational records.
For a Realtime voice agent, the selected realtime provider receives live audio plus session instructions, conversation state, and available tool definitions and returns audio and transcript events.
See Models and architecture for the current provider choices.
Knowledge and AI processing
Knowledge ingestion extracts text and sends chunks to OpenAI text-embedding-3-small to create embeddings. Retrieval sends the caller's search query for embedding, then matches it against stored vectors. Retrieved text can then be included in the configured conversational model's context.
Vox can send attachment content or image data to its configured analysis and conversational model path. Website research can also involve direct public-page fetching and a reader fallback service.
Functions and external systems
Functions can send caller-provided values, summaries, appointment details, contact fields, and tool-specific arguments to the configured destination. Tool events and results may also be stored in Wirevox logs.
Minimize function inputs and confirm the privacy behavior of every connected provider. Disconnecting an integration from Wirevox does not erase records already created in the external system.
Provider-policy boundary
Source code proves what Wirevox transmits, not every provider's current retention, residency, subprocessors, or model-training terms. Verify those properties through the applicable agreement and current provider documentation before processing regulated or highly sensitive data.