Way 🌊: A Calm AI You Can Talk To, Built with Streaming, Voice, RAG, and Real‑Time Water

Way is a contemplative AI experience inspired by Lao Tzu. You arrive, you speak or type what's on your mind, and you get a reflective response, while a living "water surface" moves behind the words like a quiet companion. Try it here: way-lovat.vercel.app.

This is the story of the tech we used to make it feel simple.

The experience we wanted

Most AI chat apps feel like forms: type → submit → wait. Way aims for something softer:

• Immediate, streaming replies (so it feels like a conversation, not a page load)
• Optional voice (so you can speak thoughts that don't want typing)
• Grounded tone (so it doesn't drift into generic "internet wisdom")
• Ambient water (so the UI feels like a place, not a tool)

Next.js, the "home" for everything

We built Way with Next.js (App Router) so the UI and server logic can live together cleanly:

• The app renders the interface and handles routing.
• Server endpoints handle chat streaming and voice-related requests.
• It deploys cleanly to production without complicated infrastructure.

Vercel AI SDK, streaming that feels alive

The biggest "feel" upgrade is streaming: the assistant's reply appears as it's being generated.

That tiny difference changes the vibe from "submit a ticket" to "someone is responding." It also keeps the app feeling fast even when the model takes a moment.

RAG, keeping the reflection anchored

A philosophical voice can easily become vague. To keep Way grounded, we use RAG (Retrieval Augmented Generation):

• We chunk the Tao Te Ching text
• Create embeddings for each chunk
• On each user message, we retrieve the most relevant passages
• Then we feed those passages as "context to reflect on," so the response stays close to the source

This is how Way can feel more like it's drawing from a real text, without turning every answer into a quotation dump.

Convex, the backend brain (including vector search)

Convex powers the backend pieces that make RAG and data workflows practical:

• Actions/mutations/queries to orchestrate ingestion + search
• Vector storage + vector search to retrieve relevant passages quickly
• A solid foundation to expand persistence later (without overbuilding upfront)

Voice, using the browser's native speech tools

Way supports voice through browser capabilities:

• Speech-to-text to capture what you say
• Text-to-speech (where supported) to speak the reflection back

The goal isn't "voice assistant" energy, it's more like speaking into still water and hearing it answer.

WebGPU + WGSL, the water isn't a video, it's real-time graphics

The Waterball background is powered by WebGPU, with shaders written in WGSL. That gives us:

• Real-time simulation/rendering on the GPU
• A visual layer that feels organic and responsive
• A calm ambient presence that supports (rather than distracts from) the conversation

Credit: Waterball inspiration (and source project)

The Waterball simulation work that inspired this integration is credited to matsuoka-601 and their open-source WaterBall project, "Real-time fluid simulation on a sphere implemented in WebGPU." See: WaterBall on GitHub.

Why this stack worked

We picked tech that supports the feeling of the product:

• Next.js for a clean full-stack app surface
• Vercel AI SDK for streaming conversation
• OpenAI + RAG for grounded, less-generic reflections
• Convex for vector search and backend workflows
• WebGPU for a living background that makes the experience memorable
• Browser speech APIs for voice that removes friction