Skip to main content
Voice AI

Mossaic: Ask your research out loud

Voice-first recall built on Moss. Product teams ask questions through a browser voice agent. Researchers ingest findings through Slack. One shared index answers both.

Mossaic

Mossaic is a reference implementation for voice interfaces on Moss, the in-process semantic search engine. It ships two surfaces that share one index: a browser-based WebRTC voice agent for product teams, and a Slack bot for researchers to upload findings.

A document ingested via Slack is answerable by voice within seconds, with no redeployment. Retrieval is grounded in source documents and returns citations alongside answers. The system handles rate limiting, session logging to Neon Postgres, and sops-encrypted secrets.

The demo corpus is a fictional business-spend product called Outlay: interviews, survey results, personas, customer success notes, and quarterly reports. It gives the voice agent a coherent dataset to reason over and makes the retrieval behavior concrete.

Key Integrations

MossSemantic Search
LiveKitWebRTC Voice
OpenClawGateway + Extensions
NeonSession Logging

Architecture

Two trust surfaces, one shared Moss index. No data duplication, no redeployment required when documents change.

Voice Flow (Public)

1
BrowserWebRTC audio capture
2
VercelNext.js web client
3
Moss runtimeManaged Python worker host
4
LiveKit workerVoice agent (Python)
5
Moss indexSemantic retrieval + citations

Admin Flow (Slack)

1
SlackResearcher uploads findings
2
OpenClawGateway with custom extensions
3
moss-knowledgeIndexes document into Moss
4
Moss indexImmediately queryable
5
Neon PostgresSession + event logging

Technology Stack

MossSemantic Search
Python LiveKitVoice Worker
Next.jsWeb Client
OpenClawGateway
Neon PostgresSession Logging
LiveKit CloudWebRTC
Slack Bot APIDocument Ingestion
WebRTCAudio Transport
sopsSecret Management

Have a system to build?

Let's scope it.