Smart Assist

Project Overview

Smarteeva’s product documentation spans thousands of pages — complex, dense, and spread across multiple versions. Internal teams and customers struggled to find accurate information quickly.

I designed Smart Assist, an AI-powered documentation search assistant integrated directly into the Docusaurus ecosystem. It enables natural-language queries, retrieves relevant content through a hybrid RAG pipeline, and returns citation-backed answers with source links.

The goal: Reduce time-to-answer, increase documentation adoption, and empower teams with fast, trustworthy self-service.

The Problem

Before Smart Assist

• Users manually searched large docs, often taking several minutes

• Keyword search returned inconsistent results

• Support teams answered repetitive low-complexity questions

• Users couldn’t find specific API or workflow details

• Release notes and feature updates were buried

Core Pain Points

• “I don’t know the exact keyword.”

• “The answer exists somewhere, but I can’t find it quickly.”

• “Documentation is too long and technical.”

• “Docusaurus search doesn’t understand context.”

Business Impact

• High support load

• Low documentation adoption

• Increased onboarding time

• Internal teams losing productivity searching for answers

Users / Audience

• Implementation specialists

• QA engineers

• Customers using the Smarteeva platform

• Support teams

• Developers and internal product stakeholders

User Needs

• Instant, accurate AI-generated answers

• Clear citations linked to real documentation

• Natural-language understanding

• Readable, scannable formatting for long answers

• Confidence that answers are grounded in real docs

• Fast navigation from answer → source page

Goals

• Reduce documentation search time by ~80%

• Improve onboarding speed for new customers

• Provide reliable, citation-based responses

• Increase documentation adoption across teams

• Reduce dependence on support

Architecture Overview

Smart Assist is powered by a hybrid RAG architecture designed for documentation that changes frequently.

1. Static Documentation → OpenAI Vector Store

Stable documentation (SOPs, long-form guides, API references) is embedded once and stored in OpenAI’s vector store.
This reduces reindexing overhead and provides fast, static retrieval.

2. Dynamic Documentation → Pinecone Vector Store

Content that changes regularly — such as release notes and versioned updates — is processed separately.

3. Automated Update Pipeline (AWS EC2)

When an admin triggers an update:

• EC2 instance fetches updated MDX pages

• Content is cleaned, chunked, and embedded

• Pinecone is updated with only the changed segments

• Reference metadata (URL, section, version) is stored

This ensures the assistant always reflects the latest documentation.

4. Unified Retrieval

User Query → Query Embedding → Parallel Retrieval

• OpenAI Vector Store (static)

• Pinecone Vector Store (dynamic)

Results are merged and deduplicated into a single Top-K context.

5. LLM Reasoning

Merged chunks are passed to an LLM (GPT-4/4o) with constraints ensuring:

• grounded answers

• no hallucinations

• accurate source citations

• readable formatting

6. Response Delivery

• Markdown output

• Citation blocks

• Reference links

• Scroll-optimized UI for long answers

UX & Interaction Design

Key UX Decisions

• Clean, minimal chat interface embedded within Docusaurus

• Two-panel layout: questions on the left, answers on the right

• High-clarity formatting for long responses

• Example prompt suggestions for discoverability

• Smooth scrolling and spacing for readability

• Trust-focused citation presentation

UI States Designed

• Empty state with onboarding prompts

• Loading skeleton

• Retrieval failure / fallback state

• Answer state with citations

• Long response scroll handling

• Error-handling patterns (timeouts, empty retrieval)

The Process

1. Research & Discovery

• Analysed documentation structure (MDX, versions)

• Reviewed search logs to understand real queries

• Shadowed support teams

• Identified patterns in repeated questions

2. Concepting & Iteration

• Explored multiple layouts and prompt-suggestion models

• Designed early dark-mode interface for clarity

• Built prototype to test chunk readability

• Iterated on citation placement for trust

3. UX Design

• Semantic search experience integrated into chat UI

• Markdown-friendly answer pane

• Smooth transitions for long responses

• Mobile-responsive layout

4. Implementation Handoff

• Provided detailed Markdown response spec

• Chunk-to-UI formatting rules

• Error and fallback patterns

• Citation grouping logic

• UI integration guidelines for Docusaurus

Impact

Quantitative

• Search time reduced from minutes to seconds

• Documentation adoption increased across 4 internal teams

• Significant reduction in support escalations

• Faster onboarding for new users

Qualitative

• “This is the fastest way to find anything in the product.”

• “Finally answers I can trust because of the citations.”

• “Much better than the default Docusaurus search.”

• “Having links back to the source page is extremely useful.”

Challenges

• Designing for long, technical answer structures

• Keeping responses grounded when context was missing

• Managing noise from multiple document versions

• Balancing minimal UI with powerful features

• Ensuring readability across long Markdown outputs

• Handling irregular documentation formatting

Reflection

This project strengthened my ability to design AI-first interfaces that balance accuracy, clarity, and trust. I learned to:

• Think in retrieval-first design

• Blend UX craft with AI system architectures

• Design UI for long-form AI responses

• Create trust through citations and reference mapping

• Work across engineering, data, and product teams

Smart Assist became one of the most impactful tools for internal productivity and customer onboarding at Smarteeva.

Flowchart

Screens