Skip to content
Anas KhanLet's talk

VoiceForge AI

Voice conversations, grounded in documents.

Project
RAG-powered voice agent platform
Year
2026
Discuss this project
Workspace overviewThe VoiceForge workspace, bringing agents, knowledge and conversations together.
A live voice conversationBrowser voice interaction with a speech waveform and conversation transcript.
AI agent workspaceAgent roles, status, configuration and available tools in one workspace.
Knowledge and retrievalDocument ingestion, indexing and a playground for testing retrieval.
Containerized servicesThe VoiceForge services running in Docker Desktop.
VoiceForge AI
VoiceForge AI workspace with voice-agent setup, knowledge shortcuts and dashboard panels
Workspace overview

The VoiceForge workspace, bringing agents, knowledge and conversations together.

1 / 5

The problem

A voice interface is only as useful as the information behind its answers. VoiceForge AI brings document retrieval into a real-time browser conversation, with source citations that let a user trace an answer back to its evidence.

Implementation

The system combines a Next.js browser experience with Python and FastAPI services. PostgreSQL and pgvector support the document knowledge layer, while hybrid retrieval and document-aware filtering shape the evidence available to a response. Docker packages the services for deployment.

Workflow overview
  1. Browser voice interaction
  2. Hybrid document retrieval
  3. Relevant evidence
  4. Response with citations

Features

  • Real-time voice conversations

    Interact with the assistant through voice directly in the browser.

  • Hybrid document retrieval

    Find supporting information using hybrid retrieval and document-aware filtering.

  • Visible source citations

    Trace a response back to the documents that supplied its evidence.

  • Expanded evidence window

    Use up to 6,000 characters of retrieved context to support a response, increased from 800.

Outcomes

retrieved evidence capacity
7.5×
800 → 6,000 characters of context
  • Expanded available evidence context from 800 to 6,000 characters.
  • Added document-aware filtering to keep answers relevant to their sources.

This project connects an interactive voice experience to a grounded retrieval workflow: a browser interface, document data, backend services, and visible source attribution.

TradeM8 trading workspace with a candlestick chart, chart controls and an asset watchlist
Next projectTradeM8