Project Update - Strengthening Our Wiki

This week marks the end of week 4 of development on our capstone project. My focus was getting the LLM connection working for a wiki summarizer service. I tested the setup both locally using LM Studio and remotely through our client’s hosted LLM endpoint. I learned a lot about how LM Studio manages local LLM instances and how containerized services communicate with them. I had to think about system resources more carefully, since running large models locally requires significant GPU memory and CPU allocation to avoid slow responses or failures during summarization. I also learned how requests and tokens work in LLM-based APIs. Each token represents a piece of text that contributes to both input and output length, affecting processing time and model cost. Understanding these limits helped me troubleshoot the incomplete responses we were receiving and adjust prompt sizes for better summarizations. To handle longer documents, I implemented a chunking mechanism that splits text into smaller sections before summarizing, allowing the model to process large inputs efficiently without exceeding token limits.

Meanwhile, my team worked on enhancing the crawler with additional filters to skip unnecessary or duplicate pages, improving the overall quality of the extracted data. They also continued developing the publisher service and website output to make sure the content is organized and displayed cleanly.

Next week, our plan is to focus on improving the UI/UX of the wiki, so the summarized content is easier to navigate and visually consistent. My job will also be to work with teammates to help extend the summarizer to handle both Traditional and Simplified Chinese. In addition, we’ll begin setting up the automated biweekly crawl. This will allow the system to refresh and summarize new content.

As for project enhancements, we will be implementing the LLM for Comparative Narrative Analysis, where the model will act as an evaluator comparing its own summaries to BART scores and human-reviewed examples. This will help us measure and compare qualities between different models.

At this stage, our team is working very smoothly together and has strong synergy. Everyone has been contributing consistently and communicating effectively. We currently have all the resources and support we need from both our faculty advisor and our client team. There aren’t any major blockers at the moment, and we feel confident continuing development on our own.