What Are Recall and Precision in eDiscovery?

Recall and precision come from information retrieval, the field behind every search engine. eDiscovery borrowed them because discovery is, at its core, a retrieval problem: find the relevant documents in a large collection without producing the irrelevant ones.

The two metrics answer different questions. Recall measures completeness. Of all the relevant documents in the collection, what percentage did the process find? Precision measures accuracy. Of the documents the process retrieved, what percentage are actually relevant?

A short example makes the difference concrete. Suppose a collection holds 100 relevant documents, and a search returns 50 documents, 40 of them relevant. Recall is 40 percent, because 40 of the 100 relevant documents were found. Precision is 80 percent, because 40 of the 50 retrieved documents are relevant.

Recall vs Precision: The Trade-Off

Recall and precision pull against each other. Broaden a search, and recall rises while precision falls, since more relevant documents surface along with more noise. Narrow it, and precision rises while recall falls, since cleaner results come at the cost of missed documents.

In eDiscovery, the two sides of that trade-off carry different weights. Recall is a defensibility question, because a missed responsive document is a production gap. Precision is a cost question, because every irrelevant document that reaches review is billable time spent on noise. Producing parties generally optimize for recall first, then use precision to control review spend.

The F1 score combines the two into a single number, the harmonic mean of recall and precision. It is standard in machine learning evaluation, though eDiscovery validation usually reports recall and precision separately, since they answer different legal questions.

How Are Recall and Precision Measured?

No one knows the true number of relevant documents in a collection, so both metrics are estimated through statistical sampling.

• A control set, a random sample coded by experts up front, provides a baseline for estimating recall and precision as review proceeds.
• An elusion test samples the discard pile, the documents marked non-relevant or never reviewed. The elusion rate is the share of that sample that turns out to be relevant, which estimates what the process missed.
• Those sample results support a recall estimate across the full collection before the review is certified complete.

Validation applies to TAR and search-term workflows alike. Parties increasingly negotiate the validation method, and sometimes a recall target, in the ESI protocol before review begins.

Why Do Recall and Precision Matter?

They are how a review process proves it worked. Courts accept technology assisted review when the process is transparent and validated, and recall is the metric that validation centers on. A documented recall estimate, built on defensible sampling, is what answers a challenge to the sufficiency of a production.

The metrics also frame the limits of search methods themselves. Research on keyword retrieval shows how much a review can miss without measurement, which is one reason meaning-based methods like concept search and active learning exist. Measuring recall and precision turns "we searched" into "we found what was there."

‍

Frequently Asked Questions

What is the difference between recall and precision?

Recall is the percentage of all relevant documents in a collection that a process finds. Precision is the percentage of retrieved documents that are actually relevant. Recall measures completeness, and precision measures accuracy.

Which matters more in eDiscovery, recall or precision?

Recall usually carries more weight, because a missed responsive document creates a production gap and a defensibility risk. Precision matters for cost, since low precision sends irrelevant documents to billable review.

What is an elusion test?

An elusion test samples the documents a review marked non-relevant or never reviewed. The share of that sample that turns out to be relevant is the elusion rate, which estimates what the process missed and supports the recall calculation.

What is the F1 score?

The F1 score is the harmonic mean of recall and precision, combining both into a single number. It is standard in machine learning evaluation, though eDiscovery validation typically reports recall and precision separately.