AI Support Agents / SaaS

WIDGENIX

Train an AI agent on your knowledge base, then answer customers everywhere — embeddable chat and real-time voice from one project dashboard, with human handoff when it matters.

View Live Platform

NestJS 11TypeScriptPostgreSQL

Scroll

Chat + Voice Channels

Shared Knowledge Base

Steps to Go Live

Always-On RAG Retrieval

The Challenge

Support tools answer from scripts. Customers need answers from your knowledge.

Most teams stitch together a chatbot, a phone tree, a help center, and a human inbox. The bot guesses. Voice is a separate product. Escalations lose context. And every new FAQ means another round of prompt-tuning. Widgenix trains one agent on PDFs, URLs, and your help content, then deploys that same brain as a chat widget and a voice widget — with RAG on every message, usage metering, and a live handoff path when AI should not guess.

The Problem

Without Widgenix

Generic chatbots, disconnected voice tools, and manual FAQ updates. Answers drift from the source of truth, billing and VIP cases get guessed, and agents re-ask questions the visitor already typed.

✓

The Solution

With Widgenix

Point the platform at your docs once. The same knowledge powers typed chat and spoken voice, with always-on retrieval, lead capture, MCP tools, and a Socket.IO handoff so your team can take over live.

Core Capabilities

One agent. Every way customers reach you.

KNOWLEDGE

Train on PDFs & URLs

Upload documents or paste page links. Ingestion chunks, embeds, and stores vectors in PostgreSQL + pgvector so the agent answers from your content — not a generic model.

How It Works

Live in three steps — then keep improving.

Connect your knowledge

Upload PDFs or add URLs. A Bull ingestion job parses, chunks, and embeds into pgvector so retrieval is ready on the first visitor message.

Configure the agent

Set the AI persona, languages, theme, allowed origins, and tools. One project owns the knowledge base, widget, and voice settings.

Deploy everywhere

Paste the embed script. Chat streams over SSE; voice uses the same RAG brain. Public visitors authenticate with a widget public key, not a user login.

Answer with RAG + tools

Every message retrieves knowledge, then a LangGraph agent can search the KB, capture leads, look up locations, call MCP tools, or request a human.

Handoff and refine

Review conversations, tighten sources, and take over live chats when the visitor needs a person — same project, same rules, on web and soon on mobile.

Connect your knowledge

Upload PDFs or add URLs. A Bull ingestion job parses, chunks, and embeds into pgvector so retrieval is ready on the first visitor message.

Configure the agent

Set the AI persona, languages, theme, allowed origins, and tools. One project owns the knowledge base, widget, and voice settings.

Deploy everywhere

Paste the embed script. Chat streams over SSE; voice uses the same RAG brain. Public visitors authenticate with a widget public key, not a user login.

Answer with RAG + tools

Every message retrieves knowledge, then a LangGraph agent can search the KB, capture leads, look up locations, call MCP tools, or request a human.

Handoff and refine

Review conversations, tighten sources, and take over live chats when the visitor needs a person — same project, same rules, on web and soon on mobile.

AI Pipeline

Always-on RAG, not a bolted-on search.

Every widget message retrieves from your knowledge base before the model replies — then LangGraph tools decide whether to search again, capture a lead, or escalate.

Multi-query retrieval

The current message plus recent history hit pgvector with cosine search, top-K, and a max-distance cutoff.

LangGraph ReAct agent

OpenRouter models stream tokens and tool events over SSE: kb_search, lead_capture, location_lookup, human_handoff, and MCP actions.

Local embeddings

Xenova MiniLM (384-d) embeds chunks on ingest so vectors stay in your Postgres — no separate vector SaaS required.

Live Widgenix chat widget answering from a trained CoachVantage knowledge base

Channels

Chat and voice, one brain.

Typed chat and spoken voice share the same project, knowledge, and persona — so answers stay consistent wherever visitors reach you.

Chat iframe + SSE

A loader script injects a floating launcher and frame. Replies stream token-by-token with history rehydration per session.

Voice in real time

ElevenLabs TTS and WebRTC signaling let visitors speak naturally, with a voice-mode prompt on the same RAG path.

Public-key security

Widgets are scoped by publicKey, allowedOrigins, and CSP frame-ancestors — rate-limited separately from dashboard users.

Widgenix voice widget trained on the same knowledge as chat

The Platform

Projects, knowledge, widgets, and ops — connected.

Every module purpose-built for a specific operational need.

Workspace

Projects

Agent

AI Personas

RAG

Knowledge Base

RAG

Vector Store

Channels

Chat Widget

Channels

Voice Widget

Ops

Conversations

Ops

Live Handoff

Growth

Leads

Integrations

MCP Tools

Platform

Billing & Quotas

Team

Project Sharing

Platform

Usage Analytics

Admin

Superadmin

Enterprise-grade AI agents with omnichannel support and guardrails

HANDOFF

Live

Bot → Human

Team Takeover

AI first. Human when it matters.

Routine questions stay with the agent. Billing, VIP, or anything the model should not guess can trigger a human_handoff tool — moving the conversation from bot to pending to a live teammate. The dashboard and Socket.IO gateway keep chat and WebRTC voice takeover on the same project rules. Mobile agent apps will put that inbox in your pocket.

Same project, same rules

Owners and accepted sharees take over from the portal. Escalations land in a team inbox instead of forcing an AI guess — and the visitor never starts over.

Platform Walkthrough

Screenshot Tour

Explore every screen in pixel-perfect detail.

Zoom

Screen 1 of 8

Widget in action

A live demo answering from a trained knowledge base — visitors type in the embed, and the agent streams structured replies from the project's docs.

Built With

Tech Stack

NestJS 11

TypeScript

PostgreSQL

pgvector

Sequelize

Redis

Bull

Socket.IO

LangChain

LangGraph

OpenRouter

ElevenLabs

Stripe

Passport JWT

Amazon SES

Docker

Schedule a Meeting
with Our Experts

Share a brief about your project and get a guaranteed response within 24 hours.

Support tools answer from scripts. Customers need answers from your knowledge.

Always-on RAG, not a bolted-on search.

Every widget message retrieves from your knowledge base before the model replies — then LangGraph tools decide whether to search again, capture a lead, or escalate.

Multi-query retrieval

The current message plus recent history hit pgvector with cosine search, top-K, and a max-distance cutoff.

LangGraph ReAct agent

OpenRouter models stream tokens and tool events over SSE: kb_search, lead_capture, location_lookup, human_handoff, and MCP actions.

Local embeddings

Xenova MiniLM (384-d) embeds chunks on ingest so vectors stay in your Postgres — no separate vector SaaS required.

Chat and voice, one brain.

Typed chat and spoken voice share the same project, knowledge, and persona — so answers stay consistent wherever visitors reach you.

Chat iframe + SSE

A loader script injects a floating launcher and frame. Replies stream token-by-token with history rehydration per session.

Voice in real time

ElevenLabs TTS and WebRTC signaling let visitors speak naturally, with a voice-mode prompt on the same RAG path.

Public-key security

Widgets are scoped by publicKey, allowedOrigins, and CSP frame-ancestors — rate-limited separately from dashboard users.

AI first. Human when it matters.

Same project, same rules

Owners and accepted sharees take over from the portal. Escalations land in a team inbox instead of forcing an AI guess — and the visitor never starts over.

WIDGENIX