Recall

What it is

If you have 100 documents containing the answer to a user query, and your retrieval system only surfaces 70 of them, your recall is 0.7. It is calculated by dividing the number of true positive hits by the sum of true positives and false negatives. Unlike precision, which checks if your results are correct, recall checks if you missed anything important. In large-scale systems, achieving high recall often requires scanning millions of vectors to ensure no relevant matches are excluded.

Why it matters

Recall determines the completeness of your AI features. If you are building a legal discovery tool or a medical diagnosis assistant, low recall means your system is silently dropping critical information, which can lead to compliance failures or dangerous misinformation. Ignoring recall often results in a system that feels accurate on the surface but fails to provide answers when the user specifically asks for comprehensive results.

In practice

You manage recall by adjusting parameters like the number of nearest neighbors retrieved from a vector database or the similarity threshold for a search. If you observe that your RAG application misses obvious documents from your knowledge base, you increase the limit of retrieved chunks to cast a wider net. You then monitor this by comparing system outputs against a ground truth dataset of known expected answers.

The tradeoff

Increasing recall typically forces a decrease in precision, as casting a wider net to find all relevant items inevitably pulls in more irrelevant noise.

Where it appears

Research summaries that use Recall, each linked to its source paper.

• Improving Knowledge Graph Question Answering Accuracy
  Constrained Entity Selection under Partial Knowledge for LLM-Based Knowledge Graph QA

  The paper introduces a constraint-based verification method that filters LLM responses against knowledge graph logic to significantly improve answer precision without sacrificing recall.

• Structuring Agentic Research with Evidence Graphs
  Structurally-bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch

  The paper introduces Crase, an agentic system that bounds research discovery within a citation graph to improve evidence grounding and search accuracy.

• Evaluating Legal Advice Accuracy in LLMs
  InsufficiencyBench: Evaluating LLM legal advice on underspecified user queries

  InsufficiencyBench measures how effectively LLMs identify missing information in legal queries instead of providing premature, potentially fabricated advice.

• ContractScrub: A Benchmark for Legal Contracts
  ContractScrub: A benchmark for final review of legal contracts

  ContractScrub is a new benchmark designed to evaluate how accurately large language models detect errors in legal documents.

• Automated Security Annotation for Business Processes
  A Hybrid LLM-Based Framework for Automated Security Annotation Generation in Business Process Models

  The paper introduces a hybrid LLM-based framework that automates the generation of SecBPMN2 security annotations from natural-language specifications to improve process model accuracy.