The developer community platform DEV is updating its feed algorithm to improve content discovery by integrating Gemini Embeddings 2 and the pgvector extension for PostgreSQL. The new system moves beyond simple recency or reaction-based sorting to implement a semantic feedback loop that ranks articles based on a user's unique interest profile. By generating a dynamic interest_embedding for each user, the platform calculates cosine similarity between user interests and article content within a unified semantic space. This approach allows the feed to surface intellectually stimulating posts while maintaining a balance with community social signals like author follows and reaction counts.

To maintain system stability and transparency, the engineering team implemented a wrapper pattern using Ai::Base and Ai::Embedding classes. This architecture allows for a comprehensive audit trail via the AiAudit model, which automatically tracks the model version, payload, latency, and token consumption for every vector generation task. The Gemini Embeddings 2 model represents content as 3,072-dimensional vectors. This upgrade enables native multimodal support, allowing the platform to future-proof its infrastructure for potential expansions into images, video, and audio analysis without requiring significant changes to the existing feed logic.

Additionally, the team is deploying a TrendDetector service to identify and label nuanced conversations that generic tags might overlook. Every 6 hours, a background process runs a Leader Clustering algorithm on high-quality posts, defined as those scoring at least 15 points above the homepage minimum. Articles are grouped based on a cosine distance threshold of 0.15. Once a cluster reaches 10 articles, the Gemini API is prompted to summarize and label the core debate. These findings are stored in the TrendMembership model to enhance UI sorting. The infrastructure powering these changes is available within the open-source Forem codebase.

The developer community platform DEV is updating its feed algorithm to improve content discovery by integrating Gemini Embeddings 2 and the pgvector extension for PostgreSQL. The new system moves beyond simple recency or reaction-based sorting to implement a semantic feedback loop that ranks articles based on a user's unique interest profile. By generating a dynamic interest_embedding for each user, the platform calculates cosine similarity between user interests and article content within a unified semantic space. This approach allows the feed to surface intellectually stimulating posts while maintaining a balance with community social signals like author follows and reaction counts.

To maintain system stability and transparency, the engineering team implemented a wrapper pattern using Ai::Base and Ai::Embedding classes. This architecture allows for a comprehensive audit trail via the AiAudit model, which automatically tracks the model version, payload, latency, and token consumption for every vector generation task. The Gemini Embeddings 2 model represents content as 3,072-dimensional vectors. This upgrade enables native multimodal support, allowing the platform to future-proof its infrastructure for potential expansions into images, video, and audio analysis without requiring significant changes to the existing feed logic.

Additionally, the team is deploying a TrendDetector service to identify and label nuanced conversations that generic tags might overlook. Every 6 hours, a background process runs a Leader Clustering algorithm on high-quality posts, defined as those scoring at least 15 points above the homepage minimum. Articles are grouped based on a cosine distance threshold of 0.15. Once a cluster reaches 10 articles, the Gemini API is prompted to summarize and label the core debate. These findings are stored in the TrendMembership model to enhance UI sorting. The infrastructure powering these changes is available within the open-source Forem codebase.