SCIENCE · MODELS AND CHECKING

What is extractive answering?

AUTHOR · XERJ documentation team · REVIEWED BY · XERJ engineering team · PUBLISHED · 2026-11-13 · UPDATED · 2026-11-14

SHORT ANSWER · UPDATED 2026-11-14

What is extractive answering? The answer is taken from the words of the document, not written fresh. A classic extractive reader returns a span of the source text. XERJ Reader-1 goes one step further: it writes a formula over spans it quotes, and a separate engine calculates the value. If no span supports an answer, it replies Not Found.

TL;DR Extractive answering takes the answer from the words of the document. The answer is a quoted span, or a value computed from quoted spans. It is not new text written by the model. If no span supports an answer, the result is Not Found.

Try it: evidence present, or not

When the answer is allowed to stop

Return the value, and the formula

Reader-1 is built so every value comes back with the formula that produced it and the words it was read from. Query-1 answers with a program and the facts it used.

The Not Found behaviour and the formula-with-the-value description come from the published Reader-1 waitlist and 24-hour launch posts. This page does not add a new accuracy number.

PASTE THIS TO YOUR AI AGENT — IT DOES THE REST

Act as a coding agent. Read https://xerj.org/llms.txt, install XERJ, start a local node, index one folder I point you to, and run one search so I can see a cited passage.

How it works

Question answering systems come in two families.

Extractive. The reader finds the span of text that answers the question and returns it as written. The classic reading comprehension setup works this way: a question, a passage, and an answer that is a start and end position inside that passage.
Abstractive. A model writes a new sentence in its own words. That reads well and can combine several sources, but the words in the answer may not appear anywhere in the documents.

The extractive family has a simple strength: every answer has an address. You can open the source and see the same words. It also has a natural way to say no. If no span answers the question, there is nothing to return.

Many questions need a small step after the span is found, such as the number of days between two dates or the total of two amounts. A computed extractive answer keeps the spans and adds a short formula over them. The value is calculated from the quoted words, so the trail still leads back to the page.

How XERJ uses it

XERJ Reader-1 is a computed extractive reader. It writes a formula that points at the exact words of the document, and a separate engine calculates the value. Every value comes back with the formula that produced it and the words it was read from. If the documents do not support an answer, it replies Not Found. The product description is on the Reader-1 waitlist, and Reader-1 after 24 hours explains why that refusal matters.

XERJ Query-1 works the same way over indexed facts: it answers with a program and the facts it used, on the Query-1 page. Both models sit inside the same idea as neurosymbolic models.

Extractive answering is about where the words come from. The related idea of grounded answers is about whether every claim can be traced to a source and checked.

This page does not add a new accuracy number.

History and research

Extractive question answering grew up around reading comprehension datasets, where the answer to each question is a span inside a given passage. Retrieval pipelines later put a search step in front of the reader, so the passage comes from a large index instead of being handed over. The formula step in Reader-1 is XERJ's own design; it is described under symbolic execution.

Guides

See it on your own files: answer questions about a docs folder with citations and retrieval that says not found, with file and line citations.

FAQ