When Deep Search Breaks: The Costly Mistakes Teams Make When Hunting Research Signal

Two months into a PDF-driven search project-during a March 2025 sprint to surface equations and tables from academic PDFs-the ingestion pipeline crashed in a way that looked trivial on paper and catastrophic in production. The model returned plausible but useless summaries, downstream ranking exploded in latency, and the "quick win" we promised product stakeholders turned into three weeks of firefighting and a frozen release. What went wrong is repeatable, cheap to make, and shockingly common in teams starting with intelligent search, deep research workflows, or any system that pretends to be a research teammate. Below is a post‑mortem written as a reverse guide: focus on the anti-patterns, the damage they cause, and actionable pivots that fix them. --- ## The Red Flag: The shiny thing that blew up the roadmap The trigger was obvious in hindsight: we chased a single feature that promised "instant expert answers" by wiring the LLM directly to the index and letting it "summarize everythi