1,000+ support conversations analyzed.

Once the chatbot was live, the conversation history became a valuable but difficult-to-use source of product feedback. Questions were repeated in different forms, individual conversations could contain several issues, and manually reviewing more than 1,000 transcripts was not a realistic way to identify recurring issues and knowledge gaps. The team needed a structured view of what users were asking, which problems appeared most often, and where the chatbot or underlying knowledge base could be improved.

I built a conversation-mining pipeline that cleans the raw chat history and uses NLP techniques to extract structured user issues from each conversation. The issues are embedded, stored in Qdrant, reduced in dimensionality and clustered into recurring topic areas using K-Means or HDBSCAN; an LLM then names the clusters and surfaces representative issues. A second stage decomposes compound questions and uses vector similarity plus LLM synthesis to remove duplicates and produce canonical FAQs. Sentiment, resolution status and tool-use analysis add signals about how the chatbot is performing, while a Streamlit dashboard lets stakeholders explore clusters and inspect the conversations behind them.

The pipeline analyzed more than 1,000 real conversations, surfaced 5–6 major topic clusters and generated 200 unique FAQs with suggestions for improving the knowledge base. Sentiment, resolution and tool-use analysis also exposed where users were becoming frustrated and where chatbot behavior needed attention. Instead of manually reading conversation logs, the team had a structured view of recurring demand and concrete material for improving chatbot and FAQ content.