Overview & Quickstart
docs/README.mdWelcome to Felona Voice — the ultra-low-latency, open-source voice agent framework powered by Joint Embedding Vectors (JEV).
30-Second Quickstart
Install the package from npm and start building in under 30 seconds with zero external API dependencies:
npm install felona-voice
Create your agent with deterministic actions and test a sub-millisecond turn:
import { createAgent } from "felona-voice";
const agent = createAgent("Concierge")
.system("You are an upscale hotel concierge.")
.action("book_table", "Book restaurant reservation", async () => "Table booked for 7 PM!")
.action("room_service", "Order fresh towels or food", async () => "Room service is on its way.")
.fallback("Sorry, I am not able to understand that. How may I assist you?");
// 1. Simulate a turn instantly (sub-millisecond):
const reply = await agent.interact("can I get clean towels?");
console.log(reply.text); // "Room service is on its way."
// 2. Export documentation as Markdown:
console.log(agent.drawMarkdown());
// 3. Or start streaming WebSocket voice server:
// agent.listen({ port: 8080 });Core Concepts at a Glance
Encodes user utterances into a semantic vector space and matches them against candidate actions via cosine similarity in ~5ms.
Actions execute pure TypeScript/JavaScript code, database queries, or tool calls. No hallucinations, no unpredictable prompt drift.
Build stateful conversational graphs using .addNode(), .addEdge(), and .invoke().
Generate diagrams that render automatically in GitHub, VS Code Markdown preview, and docs platforms with zero extra tools.