Rebuilding My Portfolio AI Assistant: Planning a Smarter RAG-Powered Experience

🛠️ https://jinleedev.vercel.app/

A behind-the-scenes look at how I’m planning to upgrade my portfolio chatbot from a prompt-only Gemini integration to a document-aware, OpenAI-powered RAG system using LangChain and FAISS.

 

 

🤖 Current Chatbot Architecture

 

On my portfolio website, I currently use a chatbot assistant to help visitors—especially recruiters—learn more about me. The assistant is powered by Google’s Gemini API, and it responds to predefined prompts using a single long initial prompt that includes:

 

• My bio and education
• Technical skills
• Project summaries
• Pre-written responses to likely questions

 

This setup is integrated into my site via an API route (/api/chat), and the frontend is built with React + Tailwind CSS, giving it a native chat feel.

 

---

 

⚠️ Limitations of the Current Approach

 

While it works well for casual Q&A, the Gemini-based solution has several shortcomings:

LimitationDescription

🔒 Static knowledge	It only knows what I prewrote in the initial prompt
📄 No document parsing	Can’t answer questions based on my actual resume or project files
🧠 No memory	Lacks multi-turn reasoning or context
🧪 No testable updates	I have to redeploy or manually edit the prompt for changes
📚 Limited to Gemini	Hard to swap out models or build on top of it

 

---

 

🔁 Planned Refactor: From Gemini → LangChain + OpenAI (RAG)

 

To overcome these issues, I’m in the process of transitioning the assistant to a Retrieval-Augmented Generation (RAG) setup. The new approach includes:

 

 

🧱 Core Components:

 

• LangChain: For chaining document loaders, embeddings, and LLMs
• FAISS: Lightweight in-browser vector DB to store document chunks
• OpenAI API (gpt-3.5-turbo): As the core reasoning engine
• Resume + Text files: Real source data for grounded answers

 

 

🔍 What RAG will enable:

FeatureDescription

📄 Document-grounded Q&A	Answers based on actual PDF/MD/TXT files
💡 Dynamic context	No need to stuff everything into one prompt
🛠️ Easily updatable	Just drop in new documents
🤖 Smarter responses	Better factuality and follow-up ability

 

---

 

⚖️ Comparison Table: Current vs Planned

FeatureGemini ChatbotLangChain RAG Assistant

Document awareness	❌ None	✅ Reads actual files
Prompt structure	🔒 Static only	🔄 Dynamic retrieval
Model	Gemini	OpenAI (gpt-3.5-turbo)
Accuracy	🎯 Medium	🎯🎯 High (fact-grounded)
Maintainability	🔧 Manual edits required	🔧 Drop-in file updates
Extendability	❌ Not modular	✅ Easily extendable

 

---

 

🛤️ My Refactor Roadmap

 

Here’s a rough outline of my planned steps:

 

1. ✅ Prototype the LangChain RAG pipeline using local files
2. 🔄 Replace Gemini API backend with a /api/chat route powered by OpenAI
3. 🧪 Test multi-document parsing (Resume.pdf, About_Project.txt, Roadmap.md)
4. 🎛️ Add toggles to switch between Gemini and RAG modes (for testing)
5. 🧼 Refactor UI if needed to improve responsiveness and loading behavior
6. 📜 Write updated documentation for API, prompt structure, and usage
7. 📝 Publish a follow-up blog with the full code and reflection

 

---

 

📝 Why This Matters

 

This upgrade isn’t just about showing off AI integration—it’s about giving my visitors (and potential employers) a real-time, accurate, and personalized experience when learning about my skills and projects.

 

By grounding answers in actual files, I remove fluff and show what I’ve truly done.

 

---

 

🚧 In Progress — Stay Tuned

 

This project is still in progress, and I’ll be documenting each milestone as I go. Once the refactor is complete, I’ll update this blog with a full walkthrough, code samples, and deployment tips.

 

If you’re building an AI-enhanced portfolio too—or thinking about it—I’d love to hear how you’re approaching it!

 

---

 

📝 블로그 한글 요약: 포트폴리오 챗봇, 더 똑똑하게 리모델링하기

 

현재 내 포트폴리오 사이트에는 Google Gemini API를 사용하는 간단한 챗봇이 있고, 방문자가 나에 대해 질문하면 사전에 작성한 정보로 대답해줘. 하지만 이 방식은 고정된 프롬프트에 의존해서, 실제 내 이력서나 프로젝트 내용을 반영하지 못한다는 단점이 있다.

 

그래서 지금은 OpenAI API와 LangChain, FAISS를 활용한 RAG(Retrieval-Augmented Generation) 구조로 개선하는 리팩토링을 진행 중이다. 이 방식은 PDF, TXT 등 실제 문서를 읽고 벡터로 저장한 뒤, 질문에 맞춰 관련 정보를 찾아 대답하는 시스템이다.

 

 

✅ 이 리팩토링의 핵심 목표:

 

• 실제 문서 기반 응답 제공 (이력서, 프로젝트 설명 등)
• 더 정확하고 유연한 질문 대응
• 쉽게 업데이트 가능한 구조
• Gemini보다 확장성과 활용성이 뛰어남

 

 

📌 앞으로의 작업 계획:

 

• LangChain + FAISS로 기본 구조 완성
• 기존 Gemini API를 OpenAI로 대체
• 멀티 문서 파싱 실험
• 필요시 챗 UI 개선
• 최종 블로그에 전체 코드와 후기 정리 예정

 

지금은 아직 개발 진행 중이고, 이 글은 리팩토링 과정과 개선 방향을 소개하는 첫 번째 블로그야. 추후 완성되면 후속 블로그로 정리할 계획이다!