🛡️ KnowledgeBot Support Assistant

An advanced, offline-first local RAG (Retrieval-Augmented Generation) customer support mobile application. (For example, configured to parse and answer queries using ConnectWise support documentation).

The system runs 100% locally on your machine—ensuring complete data privacy, offline capabilities, and zero API usage costs.

---

🎬 Application Demo

---

🎨 Color Palette Mockup Example

---

🛠️ Tech Stack & Architecture

System Components Diagram

Component Breakdown

• Frontend Client (Expo App): Built with React Native Expo (SDK 51), featuring a complete navigation stack, React Query caching, and custom markdown support. It includes a native PDF document picker that reads and uploads files as base64 streams directly inside active chat bubbles.
• Backend Server (Node.js & Express): Serves as the middle layer, hosting PostGraphile middleware to dynamically reflect database schemas as GraphQL schemas, and custom Express REST endpoints for user authentication.
• GraphQL Schema Plugin: Custom PostGraphile plugin extending the schema with high-performance askQuestion (RAG pipeline) and uploadDocument mutations.
• Vector Database (PostgreSQL 17): Core database managing tables for users, chat sessions, and messages. Employs the pgvector extension to index 768-dimension text chunk embeddings using a high-speed HNSW index optimized for L2 distance matching.
• Local AI Server (Ollama): Houses offline generative and embedding runners:
  • nomic-embed-text: Generates embeddings for document chunks and queries.
  • gemma2:2b: Generates natural, helpful customer support answers.

---

🔄 Core Workflows

A. Document Upload & Ingestion Pipeline

1. File Pick: Technicians select a local PDF file via the paperclip button on the chat input bar.
2. Base64 encoding: Client converts the file stream and calls the uploadDocument(name, base64Data) GraphQL mutation.
3. Chunking: Backend decodes the PDF, extracts text page-by-page, and splits it into semantic chunks of ~1000 characters.
4. Embedding: Backend calls Ollama's local nomic-embed-text API to convert each text chunk into a 768-dimension vector.
5. PG Indexing: Stores chunks and vectors in PostgreSQL, instantly updating the HNSW vector search index.
6. Instant Feedback: Inserts a system confirmation message bubble directly into the chat session feed.

B. RAG Querying Pipeline

1. Question: User submits a support query (e.g. "What is Backstage mode?").
2. Embed: Backend embeds the question using Ollama's local embeddings API.
3. Vector Search: Executes an L2 distance (<->) similarity query against PostgreSQL embeddings:

   SELECT content, metadata FROM embeddings ORDER BY embedding <-> $1::vector ASC LIMIT 5;

4. Prompt Engineering: Combines the aggregated top 5 context chunks, system instructions (representative persona, strict no-hallucination rules), and the user's question.
5. Inference: Local gemma2:2b processes the prompt and returns a response.
6. Output: Response is post-processed, user history is updated, citation references are resolved, and the bubble is pushed to the user.

---

🎨 Design System

The app is styled using a curated, premium color scheme:

• Primary (Blue): #29389A (Brand Identity Example)
• Secondary (Dark Navy): #0D1025 (Headlines & Dark Elements)
• Tertiary (Lime Green): #C5E654 (Accent badges & highlights)
• Neutral (Light Grey): #F6F6F6 (App-wide background)

---

🚀 Getting Started

1. Prerequisites

Install the following local dependencies:

• Node.js (v18+)
• PostgreSQL 17
• pgvector extension
• Ollama

2. Set Up Local AI Models

Open a terminal and pull the required models:

ollama pull nomic-embed-text
ollama pull gemma2:2b
ollama serve

3. Environment Variable Configuration

Server Environment (/server/.env)

Create a .env file in the /server folder matching server/.env.example:

PORT=5000
DATABASE_URL=postgresql://<DB_USER>:<DB_PASSWORD>@localhost:5432/knowledge_bot
JWT_SECRET=your_jwt_secret_key_here

Client Environment (/client/.env)

Create a .env file in the /client folder matching client/.env.example:

EXPO_PUBLIC_API_URL=http://<YOUR_MACHINE_LOCAL_IP>:5000

(Replace <YOUR_MACHINE_LOCAL_IP> with your computer's local network IPv4 address so physical devices running Expo Go can communicate).

4. Database Setup & Indexing

Run the automated initialization script:

cd server
npm install
npm run setup-db

This script enables standard extensions (vector, uuid-ossp), creates the database schema, indexes all preloaded PDFs in /docs, and computes HNSW vectors.

5. Launch Backend Server

npm run dev

6. Launch Mobile Client

Open a new terminal:

cd client
npm install
npx expo start -c

Scan the QR code in your terminal using the Expo Go app on your physical mobile device to test.

---

📜 License

This project is licensed under the MIT License - see the LICENSE file for details.