Organising project knowledge for better AI answers
Lingxi Cui, Jinhao Ruan, Wenqing Wei, Guo Yi et al. · Zhejiang University / Xiaohongshu · 14 Oct 2026 · not peer-reviewed
Source finding
EGCForest is a method that organises and links chat messages by people, chat channels and work deliverables. On the internal test, the AI produced 53.31% correct answers using this method. With the strongest comparison method, which searches for matching words and similar content, it achieved 40.70%. The same AI model answered in both cases.
Method and evidence
Comparison of retrieval and memory methods on three datasets. The displayed internal test comprises 801 questions over 10,222 messages. All methods there use GPT-4.1-mini to answer, with a 3,000-token context budget. A model grades open-ended answers. Moderate evidence for these tests.
Limitations and interests
Preprint dated 14 October 2026, not peer reviewed. One internal company dataset; country, industry and company size are not established. No measured time savings or return. Some authors work at Xiaohongshu. Model-based grading and independent replication remain limitations.
Our interpretation
Assess the organisation and provenance of project knowledge alongside the language model. These test results are not a general business success rate.
What businesses can do
Test your AI with questions about changed responsibilities and project status. Check whether it cites the correct original message, person and date.
Where to start
Table 1 and Section 4.2; Appendix A.2.3 for data and A.3/A.4 for grading. Raw values: 53.31% and 40.70%; difference 12.61 percentage points.
Connections
Data and technology · Tracing data and evidence · Reviewing project context