Chapter C05 · Interactive Course
How an answer is made
Why answers vary, and why plausible details can be invented.
Core Learning Concept
Explain an unexpected reply by tracing how it could have been generated. Separate learned language patterns from a live lookup, and separate plausible wording from evidence. Use the supplied examples to identify the likely limitation, then choose a check or repair that addresses that limitation.
What You’ll Learn in This Chapter
-
New text is not always a found answer
Distinguish creating a response from retrieving a document.
A generated answer is assembled as output; it need not be copied from a document the system just found. Retrieval is the separate act of finding existing material. An app can combine retrieval with generation, but a well-written paragraph alone does not tell you whether any source was consulted.
-
A language model works with language patterns
Recognize the role of a language model in text tasks.
A language model learns numerical patterns from language data and uses them to generate or evaluate text. Context influences what wording fits next. This can produce useful explanations and drafts, but matching language patterns does not automatically establish whether a specific date, quote, or other claim is true.
-
Text grows a piece at a time
Use a simplified sequence to explain common language-model generation.
Many language models generate text one token at a time. A token is a piece of text, sometimes a whole word and sometimes only part of one. Each chosen piece becomes part of the context for the next choice. The diagrams simplify those pieces so you can follow the sequence.
-
The same request can vary
Expect output variation without assuming one sample defines a tool.
Several next pieces can fit the same input. A generation process may choose among them, so repeating a request can produce different wording. Different context or product settings can also change a reply. Variation can help brainstorming, but it does not tell you which factual answer is correct.
-
Plausible details can be invented
Recognize fabricated names, quotes or references as possible output errors.
An invented detail can sound plausible because it resembles things that appeared in training. This is often called a hallucination. A made-up reference may look like a real citation. Check whether the source exists and supports the statement; asking for a citation is useful only if you inspect it.
-
Training information can be old
Distinguish learned information from current retrieved information.
A model’s learned information reflects the data and training it received. It is not automatically updated when a venue changes its opening hours. An app may supply current information through tools or documents. For a time-sensitive fact, check the current source and whether it was actually used.
-
Explain a surprising answer
Identify a generated rather than retrieved answer, flag an invented reference, and choose current evidence for a recent change.