Interestingly, we also observed that groups use PaperPing very differently: in some, it enhanced engagement and encouraged sharing, while in others, it soon dominated the social space, converting channels into paper feeds. This reveals the complex social dynamics of AI in group spaces. (8/9)

Social-RAG makes this work in four key steps:
1️⃣ Index past group conversations
2️⃣ Retrieve social signals (e.g., group interests) from the noisy interaction data
3️⃣ Feed them into a chain of prompts for generation
4️⃣ Post in-channel and gather continuous feedback
(5/9)

Based on Social-RAG, we built PaperPing, a Slack bot that recommends academic papers with tailored explanations that highlight connections to the group (e.g., referencing a previous relevant thread or mentioning specific group members who might find the paper interesting). (4/9)

Inspired by traditional RAG that retrieves factual knowledge for additional context, Social-RAG retrieves "social facts" from past group interactions to generate contextually relevant content aligned with group preferences. (3/9)

AI agents are entering online social spaces, but often their messages feel generic or intrusive. In our #CHI25 paper, we introduce Social-RAG, a workflow that grounds AI generations in the specific group context by retrieving from the group’s interaction history. 🧵(1/9)