Slack
Search Context
Slack’s search problem is a classic enterprise retrieval challenge: users searching their own private communication history. The product must surface the right message, file, or channel across potentially years of workspace history.
It sits at the intersection of personal information retrieval and enterprise knowledge management — more similar to email search than web search or e-commerce.
Search Challenges
- Query non-repeatability: searches are personal and contextual; no stable query distribution
- Document privacy: each user accesses a unique document set — aggregate signals don’t generalize
- Recency vs. relevance tension: the most recent message for a query is often not the right one
- Position bias in training data: enterprise users still click top results disproportionately, poisoning LTR training if uncorrected
- Work graph as signal: user interaction history (who they talk to, which channels they prioritize) is the key relevance signal
Tech Stack
- Apache Solr for first-stage retrieval (fast, configurable sorting)
- Application-layer re-ranking for second stage with richer features
- SVM via SparkML as LTR model (pairwise training on click data)
- Work graph: internal interaction history per user (author affinity, channel priority)