LLM-Powered Chatbot
A chatbot built to stay grounded — pairing GPT-4 with a vector database so answers come from a real source, not improvisation.
Overview
An independent build exploring how to make a conversational agent that knows the difference between what it knows and what it's guessing. It retrieves context from a vector database and feeds that into GPT-4, so responses are anchored to actual content.
I built a custom embeddings pipeline, experimented with prompt strategies for accuracy and tone, and added fallback logic that detects low-confidence retrievals and responds honestly instead of fabricating an answer.
Highlights
- Designed a custom embeddings and retrieval pipeline feeding relevant context into GPT-4.
- Iterated on prompt strategies to balance helpfulness, accuracy, and tone.
- Added confidence-aware fallbacks that trigger on low-similarity queries.
- Tuned chunking and top-k retrieval to keep answers relevant and concise.