Project Overview
We developed a generative AI chatbot that holds natural, context-aware conversations across a SaaS product — answering product questions, guiding onboarding, and surfacing the right docs at the right moment, all grounded in the client's own content.
The Challenge
The product had rich documentation but users could not find answers, leading to abandoned trials and a flood of "how do I" tickets. A scripted FAQ widget felt robotic and frequently sent users in circles.
- Low trial-to-paid conversion driven by onboarding friction
- Documentation was comprehensive but hard to search
- Scripted FAQ bot could not handle natural phrasing
- No personalization based on the user's plan or progress
Our Strategic Approach
We combined a large language model with retrieval-augmented generation so the chatbot answers from the client's live docs and changelog. Conversation memory and user metadata let it tailor responses to each account's context.
The Solution We Delivered
The chatbot ships as an embeddable widget with streaming responses, source citations, and a feedback loop that continuously improves retrieval quality.
- Context-aware conversations with short- and long-term memory
- RAG over docs, changelog, and support macros with citations
- Streaming token-by-token responses for instant feel
- Personalization by plan, role, and onboarding stage
- Inline feedback capture to improve answers over time
- Embeddable widget with full theming controls
Technologies Used
- OpenAI / Anthropic LLM — Natural-language understanding and generation
- LangChain — Retrieval pipelines and prompt orchestration
- pgvector — Embedding storage and similarity search
- React — Embeddable streaming chat widget
- Node.js — Streaming API and retrieval service
- Redis — Conversation memory and caching
Development Process
- Content audit — Mapped and cleaned all docs, changelog, and macros for ingestion.
- Embedding pipeline — Built an automated re-indexing pipeline triggered on content changes.
- Prompt & persona design — Crafted a helpful brand voice with strict grounding rules.
- Widget engineering — Built the streaming, themeable, embeddable front end.
- Evaluation & tuning — Ran answer-quality evals and tuned retrieval thresholds.
Results & Impact
The chatbot became the primary self-serve channel, deflecting routine questions and measurably smoothing onboarding.
- Trial-to-paid conversion improved by 19%
- 47% reduction in onboarding-related support tickets
- Median answer time under 2 seconds with citations
- Over 80% of conversations rated helpful by users
🎯 Key Takeaway
A grounded generative AI chatbot turned static documentation into an interactive guide, lifting conversion while cutting support load.
Project Gallery
