The retrieval pipeline worked. The product question remained.

In a recent technical deep dive, developer Michael Truong explores the challenges of building a retrieval-augmented generation (RAG) pipeline for a personal portfolio assistant. Initially, indexing entire documents led to poor search relevance, as individual documents often contained multiple distinct topics. To address this, Truong implemented a passage-splitting strategy based on Markdown headings, allowing for more granular similarity searches using OpenAI embeddings and pgvector. While the pipeline successfully improved retrieval accuracy, the author highlights that technical success in search does not automatically equate to a functional product experience. Testing revealed that retrieval quality is highly dependent on corpus expansion and the configuration of result limits. Truong concludes that passing automated retrieval tests is merely a starting point, and developers must carefully calibrate search settings to ensure that conversational assistants can effectively distinguish between relevant information and noise when answering user queries.