WebAgent.ai
No-code platform that lets businesses build, customize, and embed AI chatbots on their website, powered by RAG and agentic tool use.
- Built the entire platform from scratch: no-code chatbot builder, RAG pipeline, billing, and embeddable widget
- Agentic chatbot design with tool use (knowledge base search, meeting booking) instead of scripted flows
- Full RAG pipeline: website scraping, AI-powered summarization, chunking, embeddings, and similarity search
- Stripe-integrated subscription billing with tiered plans
Tech Stack
Problem & Solution
Problem: Businesses want AI chatbots on their websites but lack technical expertise. Existing solutions are either too rigid (scripted flows that break on open-ended questions) or require engineering resources to set up and maintain.
Solution: A platform where non-technical users upload content, customize the chatbot's appearance and behavior, enable tools like knowledge base search and meeting booking, and embed a production-ready AI chatbot on their website, all without writing code.
Architecture & Design Decisions
Agentic chatbot over workflow-based
The chatbot decides which tools to call (knowledge base search, meeting booking) based on the conversation rather than following a fixed decision tree. This makes it flexible enough to handle open-ended questions without rigid scripting.
Full RAG pipeline with AI-powered summarization
Web scraping → AI summarization of each page → text chunking → embedding generation → vector DB storage. At query time: embed the question → similarity search → inject relevant context → generate response. The summarization step was critical for cleaning noisy scraped content.
Embeddable widget architecture
The chatbot runs as a lightweight script tag that injects into any website without styling or JS conflicts. Users copy one line of code to go live. No iframes, no complex integration.
Supabase with RLS for multi-tenant isolation
Row-level security policies ensure each customer's data (knowledge base, conversations, settings) is strictly isolated. Combined with Supabase Auth for secure user management.
Stripe for self-serve billing
Integrated Stripe with tiered subscription plans based on usage. Customers can upgrade, downgrade, and manage billing without any manual intervention.
RAG Pipeline
The core challenge was building a system that could take any website and turn it into a queryable knowledge base automatically, without user intervention beyond pasting a URL.
Scrape
Crawl the customer's website and extract text content from every page
Summarize
Use AI to summarize each page, cleaning noisy scraped content into structured, meaningful text
Chunk & embed
Split summarized text into chunks, generate embeddings, and store in a vector database
Retrieve & respond
At query time, embed the user's question, run similarity search, and feed the most relevant chunks to the LLM as context
Agentic Chatbot Design
Rather than building a workflow-based chatbot that follows scripted decision trees, I designed the chatbot as an agent with access to tools. The LLM decides which tools to call based on the conversation context.
Knowledge base tool
When a user asks a question, the agent calls its knowledge base tool to run a RAG query: embedding the question, searching the vector DB, and returning relevant context to generate an accurate answer grounded in the business's data.
Booking tools
If the user enables meeting booking, the agent gains access to calendar tools for checking availability and scheduling meetings directly from the chat. The agent decides when to offer booking based on conversation context, not a scripted trigger.
Why agentic matters
Workflow chatbots break on unexpected inputs. An agentic design handles open-ended questions gracefully and adapts to each conversation without requiring predefined paths for every scenario.
Prompt Engineering & LLM Observability
Building a chatbot that works across different businesses with different knowledge bases required significant prompt engineering to get response quality right. The chatbot had to be accurate, stay grounded in the knowledge base, and know when to use each tool.
Iterative prompt development
Many iterations to tune chatbot behavior, ensuring it stays on topic, uses the knowledge base tool appropriately, and produces responses that feel natural rather than robotic.
Langfuse for LLM observability
Integrated Langfuse to trace every LLM call. Critical for debugging response quality, understanding tool-use decisions, and improving accuracy based on real conversations.
Results
Zero code to deploy
Non-technical users build, customize, and embed a production chatbot without writing a single line of code
Knowledge base in minutes
Paste a URL and the platform scrapes, summarizes, chunks, and embeds the entire site automatically
Agentic, not scripted
Chatbots use tools (knowledge search, booking) dynamically instead of following rigid decision trees