Building a RAG Pipeline to Reduce LLM Hallucination

Problem

At PNU x Upstage DOCUMENT AI CHALLENGE, I developed “DocDoc,” an AI service for overseas medical teams. It was a chatbot that finds relevant papers and answers questions from medical professionals.

At first, I just passed questions directly to the LLM. But problems quickly emerged.

Hallucination:

• The LLM would give plausible-sounding answers that were completely wrong
• Sometimes it even cited papers that didn’t exist
• In the medical field, wrong information can be fatal

---

Solution: RAG (Retrieval-Augmented Generation)

The idea is to search for relevant documents first and feed them as context before the LLM generates a response.

Pipeline:

1. Convert the user’s question into an embedding vector
2. Search for similar paper chunks in Pinecone (Vector DB)
3. Pass the retrieved content as context to the LLM
4. LLM generates an answer based on the context

---

Tech Stack

• Vector DB: Pinecone
• Embedding: Upstage Embeddings
• LLM: Upstage Solar Pro 2
• Framework: LangChain
• Backend: Node.js, Express

---

UX Improvement: Server-Sent Events

LLM response generation was slow. Users staring at a blank screen might leave.

Solution: EventStream (SSE)

• Stream the answer in real-time as it’s generated
• Characters appear one by one, like ChatGPT
• Users can see that a response is being generated

---

Lessons Learned

• LLMs aren’t magic. In specialized domains, hallucination is a serious issue.
• With RAG, the LLM answers based on actual documents, improving accuracy.
• Long response times can be mitigated with streaming.

---

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

• DAIC-docdoc