Is question answering the same as retrieval?

No. Retrieval finds relevant material; QA uses information to answer the question.

Must every answer be extracted word for word?

No. Systems can query structured facts, calculate, or compose explanations, provided the result remains supported.

Is “I cannot determine that from these documents” a failure?

It can be the correct response when the requested information is absent.

AI Glossary / Graphs and symbolic AI / Intermediate

Question answering

Definition

Question answering, or QA, is an AI task in which a system produces an answer to a user's question. The answer may come from documents, structured knowledge, calculations, or a model's learned information. A successful system must understand what is being asked and provide an answer that is supported within the intended context.

What it means

A search engine can return pages that might contain an answer. A question-answering system attempts to deliver the answer itself, sometimes with supporting passages or records. Questions range from short factual requests to explanations requiring several pieces of evidence.
The task is broader than a chatbot interface. A chatbot may greet people, collect information, or perform actions without answering factual questions. Conversely, a question-answering service can sit behind a search box or report tool. The central test is whether its answer addresses the question accurately, including restrictions such as a product version, date, location, or named person.

How it works

Different architectures solve the task differently. A text-based system retrieves relevant passages and extracts an answer or composes a response from them. A knowledge-based system translates the question into a structured representation and queries facts or relationships. A model can also generate an answer from learned information, although that alone does not establish current supporting evidence.
A useful design separates interpretation, evidence selection, and answer production so failures can be diagnosed. It checks whether the available material can answer the question and whether competing interpretations require clarification. The response should preserve qualifications in its evidence. For example, “available on weekdays” must not become “always available” merely because the shorter answer sounds more fluent.

Three practical examples

A help center can answer which file formats a particular product version accepts and link the relevant manual section.
An internal catalog can answer who owns a dataset by translating the question into a structured ownership lookup.
An educational assistant can answer a question about a supplied passage, distinguishing what the passage states from an inference or missing information.

Think of it this way

Think of a researcher preparing a brief answer with the relevant evidence beside it. Finding a promising page is only one step; the researcher must check that it answers the actual question. The analogy is limited because an AI system can produce convincing language without having reliable evidence or a human understanding of what its uncertainty means.

Worked example

Consider an invented documentation set. Manual A, for Cedar Editor version 2, says that exports support PDF and PNG. Manual B, for version 3, says exports support PDF, PNG, and SVG. A separate draft proposal mentions adding TIFF, but records no release that implements it.
The user asks, “Can version 2 export SVG?” A correct answer is no according to Manual A's stated complete list of supported formats. Returning “Yes, SVG is supported” from Manual B confuses versions even though the retrieved passage is relevant to the product.
Now ask, “Can Cedar Editor export SVG?” Without a version or a known current-version context, the system should distinguish versions or request clarification. It can say that the supplied version 3 manual includes SVG while version 2 does not. It should not silently assume which version the user has.
Finally ask, “When will TIFF export launch?” The proposal does not provide a confirmed launch date. A supported response identifies that gap rather than inventing a date.
For a small original test, suppose there are ten questions: seven answered correctly, two answered incorrectly, and one appropriately left unanswered because the documents lack the requested fact. Factual accuracy among the nine attempted answers is 7/9, about 77.8%. Coverage is 9/10, or 90%. These measures reveal different properties; neither alone captures whether the abstention was appropriate.

Illustrative case study

Fictional case: Juniper Guides builds a documentation assistant. Initial testing shows that it often retrieves the right product but the wrong release. The team adds release metadata to retrieval and displays the document version with every supported answer.
It includes questions with missing answers and conflicting drafts in its evaluation set. Reviewers assess whether each response addresses the requested version and whether its evidence actually supports the claim. This case is fictional and describes an evaluation approach, not measured gains or a guarantee that citations prevent errors.

Common misconceptions and mistakes

Do not equate fluent wording with factual correctness. A citation can point to a real source while failing to support the particular answer. Do not treat retrieval relevance as proof of answer accuracy, or assume a correct answer for one product edition applies to all editions.
Benchmarks also have limits. Exact matching is useful for some short-answer tasks but may reject equivalent wording or overlook broader contextual problems. Token overlap does not establish sound reasoning. A system that answers fewer questions can appear more accurate among attempted answers, so report coverage and the handling of genuinely unanswerable questions as well.

Business use

Businesses can use QA to shorten the path from a question to a supported fact. The strongest starting point is a clearly bounded collection with identifiable owners, versions, and update procedures. Choose evaluation questions from real information needs, including confusing near-matches.
Track failures separately: misunderstanding the question, retrieving unsuitable evidence, misreading the source, and generating unsupported additions require different fixes. Give users a useful route to the underlying material and a clear response when evidence is insufficient. Response speed matters, but it should be assessed alongside correctness and the effort required to verify the answer.

Check your understanding

Why is the version 3 SVG statement insufficient for answering the version 2 question?

Answer: The evidence concerns a different release. Matching the product name does not satisfy the question's version restriction.

Related terms

Learn next

Create five questions over two imaginary manuals, including one answerable only in the newer version and one with no documented answer. Write the acceptable evidence and response before testing a system. This makes incorrect version transfer and unsupported guessing visible.

Sources

Sources: Jurafsky and Martin, Speech and Language Processing, January 2023 draft, Chapter 14, question-answering paradigms and evaluation: https://web.stanford.edu/~jurafsky/slp3/old_jan23/14.pdf . Stanford Question Answering Dataset project, answerable and unanswerable reading-comprehension tasks: https://rajpurkar.github.io/SQuAD-explorer/ .

Previous: Querying Next: Questionnaire error