Gains:
- Being able to understand how semantic search finds relevant resources that keyword search misses, with its closeness in meaning, and to be able to use both methods together in place.
- Ability to critically evaluate AI-supported summaries of discovery systems according to whether they are source-based or not and confirm them with the source.
- Ability to clarify the user's vague question and distinguish 'seemingly relevant' results in terms of reliability
A user searches the catalog for “sleep problems in children,” but the top related resource is titled “Pediatric sleep disorders.” Classic search may not show this source at all because it matches the words exactly. This is where semantic search comes into play: it is a form of search that matches the meaning of words, not their spelling. In this unit, you will learn how artificial intelligence-based semantic search and discovery systems work, what they bring to the librarian, and what pitfalls they carry.
Let's define the basic concept. At the heart of semantic search is embedding: the technique of transforming the meaning of a text into a vector of numbers (a type of meaning coordinate). Texts that are close in meaning are close to each other in this number space. Thus, although "sleep problem" and "sleep disorder" are different words, they are located close together and a search for one will also find the other. This is the power of catching relevant resources that keyword search misses.
Step by step: understanding and using semantic search
1. Understand the user's true intention. Semantic search tries to capture what the user means. Your job as the librarian is to clarify the user's vague question and give the correct input to the search system.
2. Use keywords and semantic search together. Semantic search finds resources with related but different words; A keyword search is more reliable when looking for an exact term, name or code (an author name, a law number). The good librarian uses both.
3. Critically evaluate the results. Semantic search returns results that “seem relevant”; But appearing relevant does not mean being accurate or trustworthy. You evaluate the source and timeliness of the results.
4. Teach the user search strategy. Discovery systems are powerful, but users often search with single words. One task of the librarian is to teach the user to search better.
Tip: Semantic search can be weaker than keyword/domain search when looking for very precise information (a specific ISBN, a full title, all works by an author); because it confuses other results that are "close in meaning". Use domain-based search for precise information, and semantic search for discovery and topic browsing.
Discovery systems and AI-powered summaries
Modern discovery systems no longer just provide lists of results; Some produce a direct summary answer to a question with AI. This is a powerful but risky trend. It is useful if the system produces summaries based on actual sources in the collection and citing the source. But if the system generates an answer from its own general memory without citing the source, this carries the risk of hallucination and may misinform the user.
The librarian's role is to tell the user whether such summaries are source based or not. The message "The summary given by the system is a start; go to the original source and verify there" is the basis of information literacy.
Caution: Even if an AI-powered discovery summary cites a source, do not assume that the cited source actually supports that information. Sometimes the system shows the correct source but returns the wrong summary. For critical uses, open and check the relevant section of the resource together with the user.
three mini cases
Case 1 — Escaped source found. One researcher was looking for resources on the “urban heat island,” but the title of the best study in the collection was “thermal comfort in cities.” Keyword search missed this; semantic search returned it among the top five results, thanks to its semantic proximity. The researcher reached a source he would not have otherwise seen.
Case 2 — Misleading “interest.” One user searched for "economic causes of the French Revolution". Semantic search also ranked several sources generally on the subject of "revolution and economy" but unrelated to the French Revolution. The librarian realized that these "seemingly relevant" results did not actually correspond to the topic and directed the user to the correct resources.
Case 3 — Summary verification. A discovery system gave an AI summary to a question and cited two sources. The librarian opened the sources: one supported the summary, the other said the opposite of what the summary said. The system misinterpreted the resource. The librarian showed the patron the correct source and the actual finding.
Ranking, visibility and fairness
Which resource comes up at the top and which at the bottom as a result of a search is not an innocent technicality; The vast majority of users only look at the first few results. Therefore, ranking determines what information is actually seen. AI-based ranking calculates its “relevance” metric based on patterns it has learned, and these patterns may tend to highlight sources that are popular, highly cited, or in certain languages. As a result, valuable but little-known resources in new or different languages may remain invisible. The librarian's job is to know that order is not a neutral fact and to teach the user to look beyond the initial results, try different search terms, and question the order of results. The “top three sources” should never be equated with the “top three sources,” especially for a research question. Discovery requires digging deeper into the list.
Tip: When teaching a user to search, show them how to search for the same topic using two to three different sets of terms. Different terms return different result rankings; This breaks the blind spot created by a single search and opens up a richer pool of resources.
Four copyable templates
1) Search term expansion:
Enrich the search topic below for semantic and keyword searching. List synonyms, related terms, and possible formal/technical equivalents. Only suggest terms that are truly relevant to the topic. Subject: [here]
2) Clarifying the user question:
Suggest 3 clarifying questions I should ask to clarify the following vague user question (such as scope, period, genre, language). Don't assume user intent, present options. Question: [here]
3) Evaluating outcome relevance:
Below is a search topic and the result titles/summaries returned. Rate how relevant each result is to the topic as "high/medium/low" and give a one-sentence justification. Just rely on the text given. Topic: [here] / Results: [here]
4) Summary-source consistency check:
Below is a summary and the source text on which it is claimed to be based. Is every claim in the summary actually present in the source text? Mark item by item "supported / unsupported / partially". Summary: [here] / Source: [here]
Weak prompt / Strong prompt
Weak prompt:
List me the best resources on this subject.
The AI does not know from which collection, according to what criteria, to choose from actually existing resources; can produce a made-up "best of" list from his general memory.
Powerful prompt:
Your role: scout assistant. Below are the search topic and 15 actual results (title + abstract) returned from the scout system. List these 15 results according to their relevance to the topic and give a one-sentence justification for each. Adding a new source to the list, making it up. Topic: [here] / Results: [here]
The powerful prompt limits the AI to the real result set and makes it evaluate; It prevents fabrication and retrieval from general memory.
Search type comparison chart
Status
best method
Why
Exact title/ISBN/author
Field/keyword
One-to-one matching is reliable
Topic exploration, different terms
Semantic search
Captures closeness of meaning
A current event/noun
keyword + date
Accuracy and timeliness
Related but different name source
Semantic search
finds synonym
Common mistakes
- Using semantic search for precise information. Domain searching for ISBN or full title is more reliable.
- Assuming that "what seems relevant" is true. Closeness of meaning is not a guarantee of reliability or accuracy.
- Giving the AI summary without looking at the source. The summary may have misinterpreted the source.
- Not clarifying the user's vague question. Wrong input brings wrong result.
- Using only one method. It is best to use semantic and keyword search together.
In summary
Semantic search and discovery systems find relevant resources that keyword search misses by matching the meaning, not the word, and strengthen discovery. AI-powered summaries provide convenience but carry the risk of hallucination and misinterpretation of the source. The job of the librarian; Using semantic and keyword search together, critically evaluating what "seems relevant", verifying AI summaries with the source, and teaching the user a source-based, confirming search habit.
Application task
Choose a topic and do both keyword and semantic searches (if applicable); compare two sets of results: what additional resources did the semantic search find, what unrelated results did it mix up? Generate an AI-powered summary (or take a sample summary) and check whether the claims in the summary actually occur in the source with the "Summary-source consistency check" template.
checklist
- [ ] I used domain/keyword for precise information and semantic search for discovery.
- [ ] I clarified the user's vague question.
- [ ] I evaluated results that "appeared to be relevant" for reliability.
- [ ] I verified the AI summaries with the original source.
- [ ] I explained the user's source-based search habit.