Improving production rate with RAG-based content generator

Content production was the bottleneck. The team knew what it wanted to say and had the source material to say it, but turning that material into publishable posts was manual work that competed with everything else on the roadmap.

The approach

I defined the retrieval layer before touching generation. The source corpus was chunked and indexed so that every generated draft could cite the material it came from — the point was not to produce more text, it was to produce text the team could stand behind without re-reading the source each time.

Generation runs on qwen2.5b, chosen for cost and for running close to the data rather than for raw capability. The model’s job is narrow: assemble retrieved material into the house format. Anything requiring judgement stays with a human.

What shipped

A pipeline the team runs themselves — retrieval, draft generation, and a review step before anything publishes. Design, build, and deployment were all mine.

Content production went up 5x against the manual baseline, and the account added roughly 100 new followers a week over the measurement period.