Per-requirement retrieval lifted a freelance proposal tool's recall, while hybrid search and reranking hurt

A developer built GigBid, a multi-agent copilot that matches freelance job posts against a portfolio and drafts proposals citing evidence for human approval. Testing on 180 synthetic dev jobs showed that adding BM25 to dense retrieval lowered Recall@5 from 0.528 to 0.418 when the whole job post was used as the query. Splitting each post into individual requirements and retrieving for each one raised Recall@5 to 0.821 and reached 0.947 with fusion and an outcome boost. A cross-encoder reranker reduced performance, so the developer kept it only as a relevance threshold to decide whether a skill is evidenced. The author reports that the retrieval layer, not the agents, presented the largest surprises, and that results were replicated on open-weight Qwen models.