AI Learning Hub › Learn AI › Self-Paced Courses › Introduction to AI › Intermediate › Research, Search, and Grounding

Level 3 of 4 · Intermediate · Lesson 3 of 6

Research, Search, and Grounding

Framing a research question, planning sources, and separating what evidence a system used from whether its answer actually follows from that evidence.

About 20 minutes

Skip to lesson

By the end of this lesson you should be able to

  • Frame a research question with its jurisdiction or scope, its time period, and the kinds of sources that could answer it
  • Ask two separate questions of any answer: what evidence was available and how the system obtained it, and whether the answer actually follows from that evidence
  • Say what a source collection can and cannot answer, and record the next search step and open questions

Why this matters

The question for this lesson is: what evidence does this question require, and where can it come from? Answering it comes before choosing a tool. An AI research tool can help: it can help you sharpen a question, suggest search terms and kinds of source to consult, find candidate sources, and summarize what it retrieves. You still decide whether that plan fits what you are really asking, and you verify what comes back.

Start with the question and a map of sources

Before searching, write down:

  • The question, as one sentence you could hand to a librarian.
  • Scope: for legal questions, the jurisdiction and court; for others, the institution, program, or population.
  • Time period: current law or policy, or the rule as of a particular date?
  • The kind of answer: a rule, a date, a list of scholarship, a comparison.

Then sketch which kinds of source could answer it:

Primary authority
The law itself: constitutions, statutes, regulations, and court decisions. For institutional questions, the governing policy or official page plays a similar role.
Secondary sources
Material that explains or analyzes the law or a topic, such as treatises, law review articles, and practice guides. Useful for orientation and for leads to primary sources; not the rule itself.
Finding tools
Library catalogs, legal research databases, and indexes. Each covers a particular set of material, and none covers everything.
People
Librarians, and the office that owns a policy. The Robert Crown Law Library (opens in a new tab) can help you plan a search: library@law.stanford.edu.

A generated summary is a research aid, not a substitute for the authority it describes. Its statements may or may not match the underlying sources, which is why the next lesson checks them.

Two questions to ask of any answer

Tools gather and use evidence in different ways, and the ways overlap. Rather than sorting answers into types, ask two questions.

1. What evidence is available, and how did the system get it?

  • No identifiable supporting source. The answer points to nothing you can trace. It may reflect what the model learned in training, but the answer itself cannot show that.
  • Material supplied in the request. Text you pasted or uploaded. It can support a well-grounded answer without any separate search step.
  • Open-web search. The application searches the web and places some results in the context. This is one form of retrieval. It reaches what the search finds and can access; subscription databases are often out of reach.
  • Retrieval from a selected collection. The system searches a defined set, such as uploaded files, a course folder, or a vendor’s database. The evidence available is limited to that collection and to the passages the search picks out.

2. Does the answer actually follow from that evidence?

Open the cited passages. Do they say what the answer says, with the same scope? Did the answer drop a condition, or add a claim no source makes? Is a conclusion an inference whose premises need checking? An answer drawn from the web or from a collection can be well supported or poorly supported, and any generated answer can include additions that none of its sources contain. In this course, a source-grounded answer is one where this second question comes out well: its claims can be traced to passages that support them.

Retrieval-augmented generation, or RAG, is the general name for combining retrieved material with generation. The term comes from a 2020 research paper that described models combining “parametric and non-parametric memory”: what a model learned in training, plus a searchable index of documents. (Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (opens in a new tab), 2020; technical research, background only.) Retrieval adds evidence to what the model works from; it does not replace what the model learned in training, and it does not mechanically limit every sentence to the retrieved material. Products described as RAG today vary widely and are not necessarily built the way that paper describes.

What retrieval does not guarantee

  • Retrieval can be incomplete or off target. The search may miss the passage that matters or return passages that only share keywords.
  • Coverage varies. The evidence available from a collection is limited to what it covers: a collection of 2015–2020 materials is no evidence of what changed in 2023. And an answer can be incomplete even when its sources are adequate.
  • A citation can be attached to the wrong claim. The source is real; the sentence beside it says something the source does not.
  • Absence is not proof. If the sources you searched do not mention something, record that it was “not identified in the sources reviewed,” not that it does not exist.

Worked example: one question, four answers

The question and the official page are real; the four responses are constructed teaching examples, not output from any tool.

A student asks: “When was Stanford Law School’s generative AI policy last revised?” The governing source is the Office of Student Affairs page Use of Generative AI Technology (opens in a new tab). When this course checked it on September 24, 2026, the page showed no effective or revision date.

The four responses show different ways of handling evidence, not a ranking of tools.

  • Response A (no source offered): “The policy was last revised in August 2024.” Nothing is offered in support of the date. You could still investigate it independently, but the answer gives you nowhere to start.
  • Response B (cites a web page): “According to the Office of Student Affairs page [link], the policy was updated in 2025.” The link is the right page, but on the date checked the page showed no date. A real citation is attached to a claim it does not support.
  • Response C (from the student’s folder of saved course documents): “None of these documents mentions a revision date.” The answer alone does not show that; the folder would need checking before you accept it. Even if it is accurate, the folder is not the governing source.
  • Response D (cites the official page and reports the gap): “The official page [link] shows no effective or revision date. A revision date was not identified in the sources reviewed. An archived capture of the page could show when particular wording was observed, and the Office of Student Affairs could confirm the dates.”

Response D is the most useful even though it gives no date: it says what the evidence shows, where the gap is, and what to try next. Two cautions. An archived capture shows what a page said when it was captured; it can help bound when wording changed, but it does not by itself establish a revision or effective date. And the date you consulted a page is not its revision date. If an official date is later published, use it, and cite where it appears.

Try it: what can this collection answer?

A constructed scenario about a fictional law school. The handbook and its rules are invented for practice and are not Stanford policy.

A student in the fictional school’s clinic asks: “What does the clinic handbook currently require before I use an AI tool on a client matter, and when did that rule change?” A shared folder contains four items:

  • S1. Clinic Handbook, 2023–24 edition (issued by the Clinics Office, August 2023), section 4.2: “Students must not enter client information into any AI tool.”
  • S2. Clinic Handbook, 2025–26 edition (issued by the Clinics Office, August 2025), section 4.2: “Students may use AI tools approved by the Clinics Office on client matters only with the prior written approval of their supervising attorney.”
  • S3. Student newspaper article (March 2025): “Clinics are expected to relax their AI rules next year, according to a student representative.”
  • S4. A summary a classmate generated with an AI tool (undated): “The handbook bans AI in clinics.”

Coverage note: the folder holds only the 2023–24 and 2025–26 editions. The 2024–25 edition is not in it.

  1. Which items could answer each part of the question, and which could not?
  2. Which claims remain ungrounded?
  3. What can you say about when the rule changed?
  4. What is the next search step?
Answers and discussion
  1. The current requirement: S2 is the most recent official excerpt in the folder. It supports describing what that edition’s section 4.2 says: AI tools approved by the Clinics Office may be used on client matters only with the supervising attorney’s prior written approval. It does not establish that this is the complete current rule; other sections, a later edition, or a mid-year update could add requirements. The history: S1 and S2 show different wording in the two editions, and they address different things. S1 restricts entering client information into an AI tool; S2 sets conditions for using approved tools on client matters. Permission to use a tool on a matter does not by itself establish permission to upload client information, and the excerpts do not show whether a separate confidentiality or data-entry rule still applies. S3 reports a prediction, not an adopted rule. S4 is a generated, undated summary that overstates S1: S1 restricts entering client information, not all AI use. S4 is not a source.
  2. That S2 is the complete, current rule; whether S1’s restriction on client information was removed, narrowed, or moved elsewhere; and anything about what the 2024–25 edition said, whether it comes from S3’s prediction or anywhere else.
  3. Only that the wording of section 4.2 differs between the 2023–24 and 2025–26 editions. The excerpts do not establish when a change was made or took effect, what the missing edition said, or that any particular restriction ended.
  4. Obtain the full current handbook and any related provisions, such as rules on confidentiality or client information; the missing 2024–25 edition and any update records; and, where the text still leaves questions, clarification from the Clinics Office. A librarian can help locate archived editions.

End with a research handoff

Research with AI assistance should end with something the next step can use, not a polished paragraph. A short handoff keeps sources, claims, and gaps separate. Checking the claims is the subject of Lesson 4.

Research handoff

QUESTION
[One sentence, with scope and time period]

SOURCES FOUND
[Each source: what it is, who issued it, date or edition, and where you found it]

PROVISIONAL CLAIMS (not yet verified)
[Each claim, with the source and passage it seems to rest on]

OPEN QUESTIONS AND GAPS
[What the sources did not cover; what was not identified in the sources reviewed]

NEXT SEARCH STEP
[Where to look next, or whom to ask]

Know what your sources cover. Before relying on a search result or a grounded answer, ask what collection it drew on and what that collection leaves out. An answer is only as complete as the sources behind it.

Sources and further reading