How to Set Up & Deploy Saathi Bot
Everything you need to know: what settings to change, which API keys to provide, how to ingest your data, and how to embed the zero-hallucination chatbot on any website in under 5 minutes.
Setup Roadmap at a Glance
7 Easy StepsCreate Account & Initialize Your Bot
Setup time: ~30 seconds · No credit card required
Start by creating your workspace account. Each account belongs to an Organization, ensuring your team and customer conversations are strictly isolated from any other tenant.
What to fill in during creation:
- Bot Name: E.g., “Acme Support Assistant” or “Priya from Saathi”. This is displayed in the widget top bar.
- Welcome Greeting: E.g., “Namaste! How can I help you with our products today?” (Supports English, Hindi, or Hinglish).
- Suggested Prompts: 2 or 3 quick-clickable chips for visitors (e.g. “What is your refund policy?”, “Pricing plans”).
Ingest Your Knowledge Base (No Hallucination)
The bot ONLY knows what you provide here. If it is not in your documents, it will safely decline to answer.
Navigate to Dashboard > Sources. You can ingest knowledge through three different methods:
Website Crawler
Enter your docs or marketing URL. The crawler strips headers, footers, scripts, and extracts clean markdown into 500-token semantic chunks.
PDF & Word Upload
Upload PDF product manuals, warranty guides, or Word documents (.docx). Embedded using 768-dim vector embeddings directly in Supabase pgvector.
Custom Q&A / Text
Paste precise text snippets or exact Q&A pairs (e.g. “Return window: 7 days with original invoice”) for mission-critical rules.
Configure AI Model & BYOK Keys (Zero Markup)
What settings to change in Bot Settings > Model Config
Unlike legacy chatbots that charge $0.05 per conversation (a 20x markup), Saathi Bot allows you to plug in your own API key directly:
Model: gemini-1.5-flash or gemini-2.5. Extremely fast (sub-300ms), 1M context window, generous free tier.
1. Go to Google AI Studio
2. Click “Create API Key” (Free)
3. Paste key into Bot Settings > Model Config
Model: llama-3.3-70b-versatile. Blazing inference speed (under 100ms first token), open-weight intelligence.
Zero-Database Key Security
Saathi Bot enforces Zero-DB storage for client API keys. Your LLM keys are never saved to plain database tables. They are passed strictly via encrypted runtime channels, so no internal database breach can ever compromise your credentials.
Whitelist Your Website Domain (Mandatory Security)
Prevents unauthorized websites from stealing your bot or consuming your tokens
Navigate to Bot Settings > Allowed Domains. Add the exact web address where your widget will be embedded:
How Origin Verification Works:
When a browser loads the widget, the session route verifies the HTTP Origin header against this whitelist. If a malicious competitor copies your embed script to their website, their requests are automatically rejected with an HTTP 403 Forbidden.
Embed the Chatbot on Your Website
1-line script tag · Shadow DOM isolation · Zero CSS clash
Copy the lightweight embed code from your bot overview page and paste it into your website.
<script src="https://saathi-bot.vercel.app/widget.js" data-bot-id="YOUR_BOT_PUBLIC_ID" defer></script>WordPress Sites
Install any header/footer script plugin (like “WPCode” or “Insert Headers and Footers”). Paste the snippet into the Footer Scripts section and save.
Shopify Stores
Go to Online Store > Themes > Edit Code. Open theme.liquid, scroll down to the bottom, and paste the snippet right above the closing </body> tag.
Next.js / React (App Router)
In your root app/layout.tsx, import Script from 'next/script' and add:
<Script src="https://saathi-bot.vercel.app/widget.js" data-bot-id="YOUR_BOT_ID" strategy="lazyOnload" />Webflow / Wix / Squarespace
Open Project Settings > Custom Code > Footer Code. Paste the script tag and publish your site.
Customize Theme, Colors & Launcher Position
Match your company brand palette in Bot Settings > Styling
Saathi Bot renders entirely inside an encapsulated Shadow DOM. This guarantees your site's CSS (Tailwind, Bootstrap, etc.) will never break the widget styling, and the widget will never leak styles into your webpage.
- • Primary Brand Color: Select from presets or type any custom HEX color (e.g.
#0f766e). - • Color Theme: System Auto-detect, Force Light Mode, or Sleek Dark Mode.
- • Custom Avatar: Upload your company logo or support agent profile image.
- • Placement: Bottom Right (default) or Bottom Left.
- • Mobile Optimized: Smooth full-screen modal transition on smartphone viewports.
- • Bundle Weight: Ultra-fast <5KB gzip loading via edge CDN.
Test in Playground & Train Unanswered Questions
Verify accuracy before embedding live and review user questions automatically
Before sharing your widget with the world, open Dashboard > Playground:
The Playground Tester
Type questions in English or natural Hinglish. Look at the Source Inspector chip below each reply to confirm which document and page answered the question.
Unanswered Questions Tab
When visitors ask queries that don't meet the 0.78 vector similarity score, they are automatically logged in Unanswered. Click “Teach Bot” to provide the answer in one click.
Frequently Asked Troubleshooting Questions
Why does the widget show “This assistant is temporarily unavailable”?
This occurs if the bot is toggled off in settings, if the organization is suspended, or if your domain is not added to the Allowed Domains whitelist. Check your domain settings in the dashboard.
Why does the bot say it does not know the answer?
Saathi Bot enforces a strict 0.78 vector similarity threshold. If your uploaded documents do not contain relevant text for that query, the bot admits it rather than inventing fake information. Simply add that topic to your knowledge base!
Can I use both Gemini and Groq at the same time?
Yes! Saathi Bot has an automated multi-tier gateway. If your primary provider experiences high latency or rate limits, the gateway automatically falls back to your secondary provider without dropping visitor messages.
Will the widget slow down my website load speed?
Not at all. The script uses the defer or async attribute and is only ~5 KB gzipped. It executes after your main website finishes rendering, resulting in zero impact on Google Core Web Vitals (LCP/INP).