import { NextLink, Callout } from '../views/docs/prose'; ## Architecture Overview Every Guava call involves two systems running simultaneously: Guava's hosted **Dialog System** and your **Expert** — a small, long-running service that connects to Guava's API and steers the conversation. GUAVA CLOUD Dialog System audio · STT · LLM · TTS Caller audio WebSocket Your Expert Python · TypeScript · ... Your Infrastructure local or self-hosted or Guava Hosting managed by Guava ### The Dialog System The Dialog System is Guava's managed service running in the cloud. It handles everything time-sensitive during a call: receiving the caller's audio, running speech-to-text, querying the language model, synthesizing the response, and streaming it back to the caller. Because the entire pipeline runs as a fully integrated architecture rather than a chain of off-the-shelf APIs, the Dialog System delivers best-in-class latency and naturalness. Callers hear a response that feels like a real conversation, not a chatbot reading from a script. ### Your Expert Your Expert is the code you write. Using the [Guava SDK](/docs/agent) (Python or TypeScript), it connects to the Dialog System over a persistent WebSocket and steers the agent in real time — setting its persona, sending mid-call instructions, responding to events like [`on_question`](/docs/on-question) or [`on_action`](/docs/on-action-request-execute), and issuing commands like [`transfer`](/docs/transfer) or [`hangup`](/docs/hangup). Because your Expert is just code, you can do anything: call your CRM, query a database, hit an external API, or chain into another specialized AI sub-agent. For the most common patterns — intent detection, document Q&A, vector search — Guava ships a [helper library](/docs/intent-helpers) so you can get up and running fast without reinventing the wheel. Your Expert is not in the latency-critical path. The Dialog System handles all real-time audio processing independently — your Expert can spend time on complex reasoning, external API calls, or chaining multiple AI models without the caller ever noticing a pause. During development, your Expert runs on your local machine, and Guava routes calls to it directly. You can rapidly iterate by changing the code and restarting the process — no public web server or ngrok required. ### Deployment When it's time to move to production, you'll want your Expert deployed in a highly-available configuration, running continuously and ready to handle calls at any time. Because Guava Experts only make outbound connections, it's easy to run an Expert behind a NAT or firewall. We recommend running multiple instances of the same Expert. Guava round-robins new calls across connected Experts, giving you horizontal scaling and redundancy by default. If an Expert instance dies mid-call, Guava will attempt to hand the call off to another active instance — which means you should keep in-memory state to a minimum and design your Expert to be stateless where possible. You have two options for deploying your Expert: - **Your Infrastructure** — deploy to your own servers, VM, or serverless compute platform. You control the environment. - **Guava Hosting** — push your Expert with a single [`guava deploy`](/docs/cli-reference) command and Guava manages the rest. See the [Deployment guide](/docs/deployment) for a full walkthrough of both options. ### What to read next The [Quickstart](/docs/quickstart) walks you through a complete working example in minutes. Once you're comfortable with the basics, the [SDK Reference](/docs/runner) covers every callback and call command in detail. If you want to see a real-world use case before diving into reference docs, the [example walkthroughs](/docs/inbound-rag-example) show full Expert implementations for common scenarios.