The Complete RagLeap Guide — Every Feature, Explained Step by Step

How to set up and run an AI business manager that answers customers on WhatsApp, Telegram, Discord, and real phone calls — with its own memory, its own knowledge base, and an executive assistant for the owner.
I'm the founder of RagLeap, and this is the complete technical walkthrough of everything it does — written for anyone evaluating it, setting it up, or curious how a multi-channel RAG + voice AI system is actually structured under the hood.
Most AI chatbot tools do one thing: answer questions on a website widget. RagLeap is built differently — it's a full AI business manager that runs across five channels (Web, WhatsApp, Telegram, Discord, and real Voice calls via Twilio), backed by PostgreSQL + pgvector for embeddings and Neo4j for a knowledge graph layer, with a persistent memory system that survives across sessions and channels instead of resetting every conversation.
This guide covers every part of the platform in the order you'd actually set it up.
Table of Contents

Dashboard — Your Home Screen
RagLeap Manager — Your AI Executive Assistant
AI Employees — A Specialized Team, Not One Bot
Documents — Building Your AI's Knowledge Base (RAG)
Integrations — Connecting Your Existing Tools
n8n Workflows — No-Code Automation
Voice AI — Answering Real Phone Calls
WhatsApp Business Bot Setup
Telegram Bot Setup
Discord Bot Setup
Database — Cloud vs. Self-Hosted
Billing & Analytics — Credits and Costs
Persistent Memory Dashboard
Language Settings — 222+ Languages
AI Settings — Model, Temperature, Your Own API Key
Account — Plans, Team, Billing
Audit History
AI Chat — Testing Before You Deploy
Embedding the Chatbot on Your Website
FAQ

1. Dashboard — Your Home Screen Every session starts on the Dashboard. It's a status check: is everything working, and what happened while you were away? The top bar shows your current workspace name, document count, total workspace size, and a workspace switcher — relevant if you're managing multiple businesses or clients under one account. On the page:

Getting Started checklist — Upload Documents, Configure AI, Start Chatting, Deploy Chatbot. Dismissible after onboarding.
Quick stats — Documents count, Storage used, Queries Today, Current Workspace.
RAG Quality panel — retrieval benchmark scores: nDCG@k, Hit@k, MRR@k, Recall@k. This is an information-retrieval metric — it measures whether the right source document surfaces in the top-k results for a test query set, not whether the generated answer's wording is correct. Above ~0.9 is considered healthy.

1. RagLeap Manager — Your AI Executive Assistant This is the feature that differentiates RagLeap from a typical RAG chatbot. Manager AI is a private assistant scoped only to the workspace owner — never exposed to customers — reachable via Web, Telegram, WhatsApp, or a direct phone call. Capabilities:

Read access — every document (public and private), analytics, team permissions, database connections, channel config, credit balance
Write access — configure channels, connect/query databases, generate and export reports (PDF/Excel/CSV), send emails, manage team members, change AI model/settings, deploy automations
Analysis — sales, financial (P&L, cash flow), business strategy, and customer behavior analysis, run against your documents or a live connected database, in natural language

Because it runs on the same persistent memory system as everything else, a conversation started on Telegram carries over to a phone call or the web app — it's one continuous context, not five disconnected sessions per channel.

1. AI Employees — A Specialized Team, Not One Generic Bot Every workspace is seeded with a set of AI Employee roles by default — each with its own permanent memory scope and area of expertise, rather than routing every query through one generic system prompt. How the learning loop works:

Roles are seeded automatically on workspace creation
Each role's skill context is injected into the active channel handler (voice, WhatsApp, web, Telegram) based on conversational intent
learn_from_conversation() runs across every channel — real conversations write new permanent memory, not just session-scoped context
A weekly Celery beat task reinforces learned patterns across all workspaces every Sunday

Inspect roles and their memory entries at /ai-employees/.

1. Documents — Building Your AI's Knowledge Base (RAG) This is the retrieval corpus everything else depends on. Upload documents → they're chunked → embedded → indexed → retrieved at query time with citations back to source. Pipeline, in three steps:

Upload — PDF, DOCX, TXT, or a raw URL
Processing — auto-chunked, embedded (Google Gemini embedding model, 3072 dimensions), and indexed
Retrieval — queries in Chat return cited answers pulled from the indexed corpus

Supported formats: PDF, DOC, DOCX, XLSX, XLS, CSV, XSL/XSLT, PPT, PPTX, HTML, JSON, XML, RTF, ODT, ODS, ODP, EML, ZIP, SQL, TXT, MD — up to 50MB/file.
Visibility model: Public documents are retrievable by both customer-facing channels and Manager AI. Private documents are retrievable only by Manager AI — the correct place for financial data, contracts, or anything that shouldn't leak to a customer-facing bot.
Post-indexing, the same nDCG@k / Hit@k / MRR@k / Recall@k benchmark from the Dashboard runs automatically here too.

1. Integrations — Connecting Your Existing Tools Rather than requiring migration into RagLeap, Integrations connects to systems you're already running, with AI-suggested automations per channel and a developer-level automation layer for custom logic. Same Public/Private visibility model applies to connected databases here as it does to Documents.
2. n8n Workflows — No-Code Automation Direct integration with n8n — trigger workflows mid-conversation, on any channel:

"Customer requested a callback" → create a CRM lead
"Urgent order query" → post to Slack/Discord
"Appointment confirmed" → send calendar invite / SMS

Setup: connect your n8n instance at /n8n-workflows/, map trigger conditions to specific workflows. Runs identically across voice, WhatsApp, web chat, and Telegram once configured.

1. Voice AI — Answering Real Phone Calls The most technically involved channel: a Twilio number routes into a pipeline of STT (Whisper or Deepgram) → RAG/AI response → TTS (OpenAI or ElevenLabs), with total round-trip latency in the 2–3 second range in production. Call routing logic:

Owner Mode — call from your registered owner number → routes to Manager AI (full private access)
Customer BPO Mode — any other number → routes to RAG-powered customer support, backed by your documents and live database

Setup:

Add Twilio Account SID, Auth Token, Phone Number → save
Copy the generated webhook URL into Twilio Console → Phone Numbers → Voice URL
Enable "Enable Twilio PSTN" → save
Test: call from your owner number (Manager AI) vs. any other number (Customer BPO)

Language auto-detection is real-time — a caller saying "switch to Tamil" or "speak in Hindi" mid-call triggers an instant language switch, no reconnection needed. TTS choice: OpenAI TTS (default, uses your existing key) or ElevenLabs (more natural, 200+ languages, needs its own key).
A Live Voice Test Console lets you test the full pipeline in-browser before wiring up a real number, with live Response/Audio/Total/Status metrics (browser testing adds ~2s of buffering vs. a real call).

1. WhatsApp Business Bot Setup Two supported providers: ProviderBest forApprox. costTwilioReliability, documentation~$0.005–0.08/msgGupshupCost, Indian company, fast +91 approval~$0.001–0.05/msg Critical constraint: a phone number cannot run personal WhatsApp and WhatsApp Business API simultaneously. Once registered as a bot number, it's locked out of normal WhatsApp use. Use a dedicated SIM or VoIP/landline number that can receive an OTP. Twilio: get a sandbox/production WhatsApp number → copy Account SID + Auth Token into RagLeap → test connection → set webhook to https://ragleap.com/api/whatsapp/webhook/ in Twilio Console. Gupshup: create a WhatsApp app → copy API Key + App Name into RagLeap → test connection → set Callback URL to https://ragleap.com/api/whatsapp/gupshup-webhook/.
2. Telegram Bot Setup Fastest setup path, typically under two minutes:

Message @botfather on Telegram, send /newbot
Copy the returned Bot Token
Paste into RagLeap's Telegram Bot page, enable, save
Register the webhook URL with Telegram's API (single browser GET or curl -X POST)
Test by messaging your bot directly

Most common failure mode: "Message Content Intent" not enabled, or a trailing space in the copied token.

1. Discord Bot Setup

Create an application in the Discord Developer Portal, add a Bot user, copy the Bot Token
Enable "Message Content Intent" under Privileged Gateway Intents
Paste Bot Token + Application ID into RagLeap, enable, save
Generate an OAuth2 invite URL (bot scope + Send Messages / Read Message History / View Channels), invite to your server
Mention the bot or message in an accessible channel

Note: these bots often show "offline" in the member list since they're webhook-based rather than maintaining a persistent gateway connection — this is expected and doesn't affect responsiveness.

1. Database — Cloud vs. Self-Hosted FeatureCloudSelf-HostedSetup timeInstant~10 min (curl installer)Data locationRagLeap infraYour infra onlyNeo4j knowledge graphIncludedIncludedOffline capableNoYesBest forStartups, teamsRegulated industries (health, finance, legal) Self-hosted is a one-time license fee rather than a subscription — relevant if data residency/compliance is a hard requirement.
2. Billing & Analytics — Credits, API Costs, and Usage Tracks AI credits, per-provider API key configuration, and cost breakdown. Credits: 1 credit = 1 chat query on RagLeap's System AI; 0.1 credit = 1 embedding chunk. Bring your own API key and no credits are consumed — costs bill directly to your provider account instead. Per-provider cost breakdown (Google Gemini, OpenAI, Anthropic Claude) shows API calls, tokens consumed, and estimated cost, with 7/30/90-day trend views.
3. Persistent Memory Dashboard The core differentiator vs. session-based chat tools: memory here is genuinely persistent, retrieved by semantic similarity rather than keyword match. Pipeline:

Store — fact/preference saved as a memory entry (manual or automatic)
Embed — vectorized for semantic search
Retrieve — relevant memories surfaced by meaning at query time
Inject — retrieved memories added to the AI's context window

Two scopes: User-level (portable across all your workspaces) and Workspace-level (shared across your team, e.g. escalation rules, pricing policy). Auto-compression kicks in when an entry exceeds a configurable token threshold. Full CRUD from this page, plus JSONL export and bulk purge.

1. Language Settings — 222+ Languages, Auto-Detected Documents are scanned for language distribution (token-weighted). Chat Language Preference supports 222+ built-in languages plus any custom ISO code (up to 2 total). Auto-detect query language, when enabled, matches the customer's actual input language; disabled, it locks to your preferred language regardless of input. This config is genuinely global — applies identically across Web, WhatsApp, Telegram, Discord, and Voice.
2. AI Settings — Model, Temperature, and Bring Your Own API Key Model choice: Google Gemini (default, fast), OpenAI GPT (complex analysis), Anthropic Claude (detailed reasoning). BYOK isn't limited to the three named providers — also supports OpenRouter, Deepseek, Mistral, and a fully Custom API mode for any OpenAI-compatible endpoint: self-hosted models, Ollama, vLLM, or any third-party inference API reachable by URL. Tuning: Temperature (0.0–1.0; 0.2–0.4 recommended for factual RAG), Top-K Results (chunks reviewed pre-answer; 3–8 typical). A Fallback Key Pool supports multiple keys with automatic failover on rate limits.
3. Account — Plans, Team, and Billing PlanPriceWorkspacesKey featuresStarter$29/mo1Web + AI ManagerPro$99/mo2+ WhatsApp + TelegramBusiness$199/mo3+ Voice, Team members, no brandingEnterprise$499/mo5+ SLA, dedicated support, custom integrations Also free on your own server — single workspace, web embed, AI Manager with RagLeap branding, no subscription. Team Members (Business/Enterprise) support per-member permission control and group-based bulk permissions.
4. Audit History Every config/settings change is logged: timestamp, workspace, action type, description, and actor. Filterable by workspace, action type, time range. Necessary once more than one person has admin access.
5. AI Chat — Testing Before You Deploy Exactly what a customer sees — cited answers, suggested follow-ups generated from your document set. Follows whatever Language Settings config is active, no separate configuration needed. Team Chat, separately, is a plain internal message board with no AI involvement.
6. Embedding the Chatbot on Your Website Generates embed code in three modes: Widget Bubble (floating button, customizable position/shape/color/size), Fixed Side Panel (pinned left/right), and Full Page / Inline iframe. Branding shown on free/trial tiers, removable on paid plans.
7. FAQ Setup time? Under 15 minutes for a working AI — documents, one channel, test in AI Chat, go live. Voice/Memory/n8n layer on afterward. Does it replace human support entirely? No — it handles the high-volume, predictable question set (pricing, hours, order status) across every channel in 222+ languages. Judgment-based or emotionally charged conversations still route to a human. Can I run my own model? Yes — any OpenAI-compatible endpoint via Custom API, including self-hosted. Data privacy? Private documents/databases are visible only to Manager AI, never customer-facing bots. Self-hosted keeps everything on your own infrastructure.

Originally published on the RagLeap blog, with product screenshots for each section. Try it at ragleap.com.