Alden Do Rosario

We used Cohere multi-modal embeddings to select the images for the articles. We found that this works better than GPT-Vision to select high-quality images that are contextually relevant to the surrounding text. This was a hard problem to solve, and Cohere's multi-modal embedding quality helped improve the quality of the selected images.

Serper took away the headache of incorporating search data, at an affordable cost.

We love the OpenAI APIs. To make this curated agent product better-than-human-level articles we used a mixture-of-models for each specific sub-task (like query planning, outline generation, content block generation, citation linking, etc). Being able to quickly benchmark and test various models for each sub-task in the generation process yielded higher quality outputs -- helping us differentiate from generic AI writers.

Excellent serverless vectorDB - we used it in our visual search engine that select awesome images for the articles.