hamza iqbal

LLM evaluation · public code

When retrieval helps—and when it hurts

An evaluation harness for asking whether retrieval improves the answer, and what it costs to get there.

Research implementation
Controlled evidence conditions feed the same reader; evaluation separates answer accuracy, retrieval quality and inference cost.
Evaluation workflow with controlled evidence conditions.

The question

Adding a retriever to an LLM introduces another place for a system to succeed or fail. A good answer can hide poor retrieval, and relevant evidence does not guarantee a correct answer.

What I built

I developed a multi-model harness for legal and general question answering, with controlled evidence conditions and per-question traces. It separates the evidence a reader receives from the quality of the answer it produces.

  • Compare no context, retrieved context, gold evidence and reranked evidence under matched evaluation conditions.
  • Inspect answer quality alongside retrieval metrics, model calls, tokens and latency.
  • Train and evaluate evidence rerankers under matched conditions.

Why it matters

The useful output is a diagnosis: whether the bottleneck is finding evidence, selecting it, or using it. That makes model and retrieval choices easier to justify.

← All projects