Building a RAG Pipeline to Reduce LLM Hallucination

Problem

Participated in “PNU x Upstage Document AI Challenge 2025” and developed “DocDoc,” a medical AI assistant for overseas medical teams.

The core feature was “paper search and summarization.” Initially, I just passed user questions directly to the LLM. The problem was the LLM kept hallucinating - making up research results that didn’t exist.

---

Solution: RAG (Retrieval-Augmented Generation)

Instead of relying solely on the LLM’s knowledge, I made it answer based on real documents.

Pipeline:

1. Pre-process medical papers (split text into chunks)
2. Convert each chunk to vector embeddings using Upstage Embedding API
3. Store in Pinecone (vector database)
4. When user asks a question, convert it to embedding and search similar chunks
5. Pass retrieved chunks + question to LLM, generate answer

---

Tech Stack

• Vector DB: Pinecone
• Embedding: Upstage Embedding API
• LLM: Upstage Solar
• Framework: LangChain
• Backend: Node.js + Express

---

Results

After applying RAG:

• Answers were based on actual papers, not hallucinated
• LLM could cite which paper the information came from
• Response quality improved a lot

---

Lessons Learned

• LLM by itself has knowledge limits and hallucination risks
• Combining LLM with document retrieval (RAG) maintains reliability
• Vector DB is a key component in AI applications

---

From developing “DocDoc,” which advanced to finals at PNU x Upstage DOCUMENT AI CHALLENGE 2025.

• DAIC-docdoc