Information Storage and Retrieval: AI and LLMs for Information Access

Lesson overview

Introduce AI and Large Language Models in information access, including embeddings, semantic search, retrieval-augmented generation, hallucination risk and evaluation.

CourseInformation Storage and Retrieval

Topic strandAI-assisted retrieval

Assessment styleWorked scenario, applied task and digital exam practice

EvidenceDefinitions, representation, method, result and interpretation

Starter

Write down the user need, the data being stored or searched, and the decision the system has to support. Then predict which representation, index, query method or governance control will matter most in this lesson.

Learning objectives

Explain how embeddings can support semantic retrieval.
Describe retrieval-augmented generation and why grounding matters.
Evaluate LLM risks such as hallucination, bias and opaque ranking.

Learning outcomes

Students can explain how embeddings can support semantic retrieval.
Students can describe retrieval-augmented generation and why grounding matters.
Students can evaluate LLM risks such as hallucination, bias and opaque ranking.

Key vocabulary, acronyms and terminology

AI: Artificial Intelligence: computing techniques that perform tasks associated with human-like reasoning, prediction, language or perception.
LLM: Large Language Model: a model trained on large text collections to generate and transform language.
embedding: Vector representation intended to capture semantic similarity.
semantic search: Retrieval based on meaning-like similarity rather than exact terms alone.
RAG: Retrieval-Augmented Generation: generating answers using retrieved sources.
hallucination: A generated answer that is unsupported or false.
grounding: Tying generated output to retrieved evidence.

Detailed teaching notes

Core concept

AI can improve information access by mapping queries and documents into semantic representations, summarising results and supporting natural-language interaction. It does not remove the need for retrieval, evidence and evaluation.

Representation choice

Embeddings represent text as vectors. Similar vectors may indicate related meaning, enabling semantic search for paraphrases, but vector similarity is not the same as truth.

Method and reasoning

Use retrieval to select evidence, then use an LLM to summarise or answer with citations. Evaluate both retrieval quality and generation faithfulness.

Risk and limitation

LLMs can hallucinate, overstate confidence and reproduce bias. A RAG system still fails if retrieval misses the right evidence or generation ignores it.

Degree-level deep dive

AI does not replace retrieval

Large Language Models can generate fluent answers, but reliable information access still needs source retrieval, grounding, citations and evaluation. The model is not a database, and fluent generation should be treated as a claim that needs evidence.

Embeddings are similarity evidence

Embeddings can retrieve paraphrases and semantically related text, but vector similarity is not truth. Keyword and semantic retrieval often need to be combined and evaluated together.

What excellent work shows

A strong answer separates indexing, retrieval and generation. It should explain Retrieval-Augmented Generation, name hallucination controls, require evidence, and state how generated claims are checked against sources.

Concrete example to study

RAG answer pipeline

Question: What changed in the assessment policy?
1. Retrieve policy v3 and policy v4 passages
2. Give passages to the LLM with citation requirement
3. Generate answer
4. Reject claims not supported by retrieved passages

Reasoning

Retrieval supplies the evidence.
The LLM summarises and explains, but should not invent policy details.
Grounding and citation checks reduce hallucination risk.

Reveal takeaway

A reliable AI access system separates source retrieval from answer generation, then checks whether the generated claims are supported by the retrieved evidence.

Worked example

Scenario

A student asks, 'Which policy documents changed the assessment rules?' in a document portal.

Worked solution

Use retrieval to find policy documents, versions and change notes.
Use embeddings to catch paraphrases such as evaluation rules or marking policy.
Give the LLM only retrieved evidence and ask for an answer with citations.
Check whether every generated claim is supported by retrieved sources.

Reveal model result

RAG is useful because the answer is grounded in documents. Without retrieval evidence, the LLM may produce a fluent but unsupported response.

\text{Define the data, choose the representation, apply the method, interpret the result.}

Applied retrieval task

Design a RAG feature for a course handbook.

Deliverables

Choose what documents enter the retrieval index.
Decide how embeddings and keyword search will be combined.
Write one hallucination check for generated answers.

Success checks

The design separates retrieval from generation.
The answer includes evidence and limitations.

Common misconception

An LLM is not a database. It can support access, but reliable answers need retrieval, grounding and evaluation.

Quick checks

1. In this lesson, why does AI matter?

Artificial Intelligence: computing techniques that perform tasks associated with human-like reasoning, prediction, language or perception.It replaces the need to understand user tasks or data representation.It is only a presentational label and has no effect on retrieval quality.

2. Which answer best shows degree-level understanding of ai and llms for information access?

Name as many technologies as possible without explaining the scenario.Give a final recommendation without showing how the data or query was represented.Define LLM, apply it to the scenario, then explain the trade-off or risk.

Digital exam practice

Example exam task

A company wants a chatbot to answer staff questions from internal policies, but managers are worried about hallucination. In your answer, define the relevant objects or data structures, use course-specific vocabulary, show the method rather than only the result, and finish with a decision about the storage or retrieval system.

Notation and technical toolkit

embedding(d) in R^ndocument vector representation Use to explain semantic similarity.

sim(q,d)similarity between query and document Use to rank embedding candidates.

RAG = retrieve + generategrounded answer pattern Use to separate evidence retrieval from answer generation.

What a good answer is expected to show

A strong answer for AI and LLMs for Information Access the answer should explain embeddings, LLMs, RAG, grounding, citation and hallucination controls. It should connect the formal or technical representation to the user's information need instead of listing terms without using them.

How to solve it

Describe how policy documents are indexed for keyword and semantic retrieval.
Use embeddings to retrieve semantically related passages.
Pass retrieved evidence to the LLM for generation.
Require citations and reject unsupported claims.

Model answer

Reveal model answer

The chatbot should not rely on the LLM's general memory. Policy documents should be indexed with both keyword fields and embeddings so the system can retrieve relevant passages for a staff query. In a RAG design, the LLM receives those passages and generates an answer grounded in them, ideally with citations. Hallucination risk is reduced by limiting answers to retrieved evidence, checking unsupported claims and offering the source documents for verification.

Practise next

Explain why semantic similarity can retrieve useful paraphrases.
Give one failure mode where retrieval is correct but generation is still wrong.

Self-marking criteria

Defines the scenario objects, data forms or system components before solving.
Uses the lesson vocabulary accurately and in context.
Shows a clear method with enough working for a marker to follow.
Connects the result back to retrieval, storage, analytics or governance.
States a limitation, trade-off or quality risk rather than presenting the answer as absolute.

Extension

Compare lexical BM25 retrieval with embedding retrieval for one ambiguous query.

Study route

Save a short worked answer from this lesson using this pattern: define the need or data, choose the representation, show the method, state the result and interpret the implication for the system.

Next lesson: Search Engines and Publishing Archives.