🧠 Personal RAG Database & Chatbot Engine — Kankatala Ganesh Giridhar

Production knowledge base powering a live, highly-personalized AI Digital Twin at brandofganesh.vercel.app

This repository houses the structured personal database and high-performance pipeline driving Ganesh's AI Digital Twin. It is optimized for retrieval-augmented generation (RAG), context-aware personalization, and edge-optimized serverless execution.

---

⚡ Serverless RAG Architecture

The chatbot is built entirely around an on-demand, zero-cost Serverless RAG design. Instead of paying for a 24/7 running server and a heavy vector database subscription, the backend runs instantly within Vercel's serverless function constraints:

┌────────────────────────────┐
│    Portfolio Frontend      │
│  brandofganesh.vercel.app  │ (HTML / CSS / JS Chat Widget)
└────────────┬───────────────┘
             │ POST /api/chat { message, history, visitorInfo }
             ▼
┌────────────────────────────┐
│  Vercel Serverless Function │
│       api/chat.js          │
│  ┌──────────────────────┐  │
│  │ 1. Embed query       │  │ (gemini-embedding-2)
│  │ 2. Cosine similarity │  │ (in-memory search over vectors.json)
│  │ 3. Top-8 retrieval   │  │ (RAG context compilation)
│  │ 4. Gemini 2.5 Flash  │  │ (SSE streaming response)
│  └──────────────────────┘  │
│                            │
│  data/vectors.json (24.8MB)│ (399 chunks pre-embedded)
└────────────────────────────┘

Why Serverless RAG?

• Zero Infrastructure Cost: Scaled down to $0/month. You only pay for active execution time, making it ideal for portfolio chatbots.
• In-Memory Speed: The pre-embedded database (vectors.json) loads directly into the serverless container's RAM. Cosine similarity calculations happen in sub-milliseconds with no external database network lag.
• Stream Delivery: Utilizes Server-Sent Events (SSE) to stream the response back token-by-token for a premium, real-time typing interface.

---

🤖 Advanced Chatbot Features

Your digital twin features state-of-the-art personalization, UX resilience, and secure routing:

1. ⚡ Onboarding & Role-Aware Greetings

• The widget collects the visitor’s Name and Role (Recruiter, Developer, Student).
• The profile is passed dynamically to the Vercel backend and appended to the Gemini system instructions.
• The AI Twin adapts its tone and opening pitch specifically to the visitor:
  • Recruiters: Warm and professional, highlighting project metrics, commercial impact, and availability.
  • Developers: Peer-to-peer technical deep-dives about compiler logic, offline sync queues, and database schemas.
  • Students: Helpful mentorship advice about hackathons, learning tracks, and campus life.

2. ⚡ Smart Quick-Replies (Onboarding Skip)

• High-intent action chips (🚀 See Projects, 📞 Contact Ganesh, 💼 View LinkedIn) are embedded on both the onboarding and welcome screens.
• Clicking any chip instantly skips form-filling, logs the visitor, and fires the request immediately for zero-friction access.

3. 🔗 Smart Same-Window Redirection

• All markdown links in the chat are inspected dynamically.
• Same-site links (like #about, work.html, or local hashes) load seamlessly in the existing tab (target="_self"), preserving user experience.
• External links (GitHub repos, LinkedIn profiles) open safely in a new tab (target="_blank").

4. 🛡️ Resilient AI Limit Interceptor

• If the Gemini API experiences a 429 quota rate limit or network timeout, the client intercepts the failure gracefully.
• Instead of displaying standard console logs or red errors, it outputs a charming, witty AI message redirecting the visitor smoothly to alternate contact pages.

---

📊 Database & Knowledge Graph Stats

The knowledge base compiles 35 source documents (including your full resume, hackathon entries, engineering manifesto, and 10 LinkedIn data CSVs) into optimized structures.

1. Chunk Pipeline Statistics

Running build_chunks.py processes raw data into semantic chunks:

Metric	Value	Description
Total Chunks	399	Complete retrieval nodes
Child Chunks	344	Fine-grained (200–400 tokens)
Parent Chunks	40	Macro context (512–1024 tokens)
Hypothetical Q&As	15	AI intent prediction pairs
Total Tokens	58,861	Dense semantic catalog
Total Categories	17	Categorized RAG routes (identity, projects, linkedin, timeline)

2. Knowledge Graph Mappings

We construct a directed Graph Database (entities.json + relationships.json) mapping your professional universe:

Metric	Count	Description
Total Nodes (Entities)	109	Skills, projects, roles, certifications, organizations, hackathons
Total Edges (Relationships)	110	Directed linkages (BUILT, USED_IN, WORKED_AT, KNOWS)

---

🛠️ Data Pipeline Pipeline Scripts

Written in Python 3 with zero external dependencies for maximum portability:

# 1. Validate data schemas, date formats, and check for placeholders
python scripts/validate_schema.py

# 2. Build child, parent, and hypothetical chunks with SHA-256 integrity checks
python scripts/build_chunks.py

# 3. Extract entities and relationships to compile the Knowledge Graph
python scripts/build_graph.py

# 4. Generate comprehensive database statistics and heatmap
python scripts/stats.py

# 5. Embed chunks to compile vectors (requires GEMINI_API_KEY)
python scripts/generate_embeddings.py --api-key YOUR_API_KEY

---

🎨 Technology Stack

• LLM Engine: Google Gemini 2.5 Flash
• Vector Models: Google Gemini Embedding (gemini-embedding-2)
• Similarity Metric: Cosine Similarity (In-Memory execution)
• API Delivery: Serverless Function (Vercel Node.js Edge)
• Real-time UX: Server-Sent Events (SSE stream protocol)
• Client Widget: Self-contained Vanilla HTML/CSS/JS

---

Built with 🧠 by Kankatala Ganesh Giridhar
The Linear Paradigm — consistent daily execution over manufactured turning points.