Vector search can understand different wording. It still needs you to decide which search boundaries are real.

A recruiter searches for a Sales Engineer. The role is clear enough: someone who can understand a customer’s technical problem, shape a solution, run demos or proof-of-concepts, and work closely with sales through the buying process. AI Search understands the brief. Then the search gets “cleaned up”: current title Sales Engineer, five years of experience, SaaS background, one specific tool from the job description, and a narrow location. The result is twelve candidates. All relevant. All easy to explain.

It feels like the search improved, but there is another possibility: it simply became smaller. That distinction matters much more now that candidate search is moving from literal keyword matching toward semantic and vector search. Semantic search can recognize related meaning expressed through different words, but it cannot know that a requirement you entered as fixed was actually negotiable. The question is no longer only “Did we use the right words?” It is also “Did we put the right boundaries around the search?”

That is where precision and discovery become useful. They are not competing sourcing philosophies. They are two different passes over the same market, designed to answer different questions.

Candidate search has changed from a wording problem to a boundary problem

Traditional keyword search has an obvious weakness: the recruiter needs to predict the words the candidate used. Search for “fintech developer”, for example, and a profile describing years of work on payment systems or banking software may look highly relevant to a human while sharing little literal vocabulary with the query. Semantic search approaches the problem differently. At a high level, vector search represents a query and professional information in a numerical space that captures relationships in meaning, so semantically related concepts can be matched even when they are written differently.

That is how Wandify describes its AI Search: the system analyzes the meaning of a query and can surface relevant profiles that use different wording. Recruiters can then add structured criteria and filters, creating a hybrid search that combines semantic relevance with explicit constraints. Wandify’s AI Search practical guide explains this workflow in more detail.

LinkedIn is moving in the same direction. Its current AI-Assisted Search materials describe a system that goes beyond standard filters and profile keywords, using context to identify qualifications that may not be explicitly listed on a résumé. LinkedIn AI-Assisted Search positions this as part of the shift from literal matching toward contextual candidate discovery.

A peer-reviewed study in Information Sciences also compared semantic similarity with a keyword-based baseline across recruitment data. In its simulated Software Engineer test, the semantic method produced an average similarity score of 0.74 versus 0.35 for the keyword method. The study reported similar advantages in several real-world datasets when terminology differed between résumés and job descriptions. These figures belong to that specific experimental framework, not to every semantic search product, but they illustrate why meaning-based retrieval has become important in recruitment search. Read the study

So one old problem becomes smaller: candidates no longer need to use exactly the vocabulary you predicted. But another problem remains. The search still needs to know what you consider mandatory.

What vector search can solve, and what it cannot

Imagine you need someone with experience designing event-driven backend systems. A semantic search can potentially recognize relevant evidence across phrases such as event streaming, asynchronous architecture, distributed messaging, Kafka-based pipelines and message-driven systems. You do not need to predict every possible phrase manually.

Now imagine the vacancy mentions Kafka. There are two very different interpretations. In the first, production Kafka experience is genuinely mandatory because the person will own a Kafka-heavy environment immediately. In the second, Kafka happens to be the tool the current team uses, while strong experience with another event-streaming technology could demonstrate the same underlying capability. Semantic search cannot decide which interpretation reflects the hiring manager’s actual intent. That is a sourcing decision.

The same issue appears with familiar criteria such as years of experience, industry background, exact titles, certifications and locations. Five years of experience may be a genuine requirement or a rough proxy for expected complexity. Fintech experience may be essential because of regulation, or simply the closest industry the hiring manager knows. A location can be fully fixed, while a preferred tool can be flexible. The goal is not to make every search broader. The goal is to know which type of criterion you are dealing with.

Use three types of search criteria

1. Hard constraints

These are genuinely non-negotiable. Examples can include legal right to work, a mandatory professional certification, a required language when the work depends on it, fixed on-site presence, security clearance, or a technology that truly must be used from day one. If the company would reject an otherwise excellent candidate because the criterion is absent, it belongs here. These constraints normally stay in both searches.

2. Core capability signals

These define whether the candidate can actually perform the work. For a Sales Engineer, that might include technical discovery with customers, translating business problems into technical solutions, demos or proof-of-concepts, solution design, and working alongside sales during complex buying cycles. The important part is the capability, not one exact phrase used to describe it. This is where semantic search is especially useful.

3. Search hypotheses

These are useful signals that may correlate with a strong candidate but are not necessarily requirements: one exact title, a preferred tool, a specific industry, a fixed number of years, experience at a certain type of company, or one familiar career path. Hypotheses are not bad criteria. The mistake is letting them quietly become hard constraints without testing what they remove from the market.

For a deeper framework on separating real requirements from the rest of a job description, see How to Turn a Job Description Into a Sourcing Brief.

Why filters still matter when search is semantic

It is tempting to think semantic search makes the precision-versus-coverage problem disappear. It does not. Semantic retrieval improves the system’s ability to understand relevance, while structured filters still determine which parts of that relevant market remain eligible.

LinkedIn published a detailed engineering explanation of this problem in June 2026. Its Hiring Assistant uses semantic embeddings to retrieve candidates based on qualification fit alongside several other retrieval strategies, and structured attribute filters are also applied inside the pipeline. LinkedIn’s engineering team describes the challenge as a balance between match quality and candidate coverage, and explains how approximate retrieval and post-filtering can compound recall losses when potentially relevant candidates are removed by structured restrictions. LinkedIn Engineering: Semantic Search for Hiring Assistant. That architecture is LinkedIn’s, not Wandify’s, so the implementation should not be generalized between products. The sourcing lesson is still useful: better semantic understanding does not make restrictive criteria free.

A search system can understand that two candidates have highly related experience while still respecting the filters the recruiter chose to apply. This is exactly why precision and discovery should be treated as separate search objectives.

Precision and discovery answer different questions

A precision search is designed for confidence. A discovery search is designed for coverage. Neither is the “correct” search, because they are built to tell you different things. A precision pass gives you a fast answer to “Can I find strong candidates inside the profile we currently expect?” A discovery pass tests “What happens when I stop treating our expected profile as the only valid path to the role?”

The Sales Engineer test

Consider a company looking for a Sales Engineer. The actual work requires someone who can understand technical customer requirements, design or explain a solution, run technical demos or proof-of-concepts, support complex sales conversations, and work between customers, sales and product or engineering teams.

Pass 1: Precision

Keep the role definition tight. The search may prioritize titles such as Sales Engineer, Senior Sales Engineer and Pre-Sales Engineer, while retaining genuine hard constraints. If German is essential because the person will work directly with German-speaking customers, keep it. If a specific location is genuinely required because the role is office-based, keep it. If the person absolutely needs hands-on experience with one technology from the first day, keep it. The purpose of this pass is straightforward: find people who fit the role with relatively little interpretation.

Pass 2: Discovery

Do not change the work. Change the assumptions around how that work might appear on a profile. The title family might now include Solutions Engineer, Solutions Consultant, Pre-Sales Consultant, and Solutions Architect where the profile contains real pre-sales and customer-facing evidence. Keep the same core capabilities and genuine hard constraints, but remove criteria that were only proxies. A particular CRM can become supporting context. An exact number of years can become evidence of seniority rather than a hard border. A SaaS-only filter can be tested against candidates from other technically complex B2B environments.

The discovery search will probably contain more near-misses. That is expected. Its purpose is not to produce the cleanest result page. Its purpose is to show you the part of the market your original assumptions could not see.

Do not merge the two searches too early

One common response is to put every possible title, skill, alternative and keyword into one enormous query. That removes most of the value of the exercise because you lose the ability to understand why a candidate appeared. Run the two versions separately, then compare three groups.

Group A: Candidates found in both searches. These are your highest-confidence profiles. Different search assumptions still led you to the same people.

Group B: Precision-only candidates. Look at why they disappeared from discovery. Sometimes this simply reflects ranking differences. Sometimes a criterion used in the precision pass was genuinely useful and deserves to stay.

Group C: Discovery-only candidates. This is the important group. Do not assume they are stronger simply because they are different. Review them and ask why the precision version excluded them. If the answer is a different title, an equivalent tool, an adjacent industry, or a slightly different career path, you have learned something about the search. If the answer is that they cannot actually do the core work, discovery has also done its job by showing you where the boundary should remain.

Treat the second search as an experiment

The goal is not to prove that broader search is better. The goal is to test an assumption. A useful workflow is to freeze the role outcome, label every major criterion, run a precision pass, run a discovery pass, compare candidate quality rather than only result counts, and then keep what the market taught you.

1. Freeze the role outcome. Write one sentence describing what the person actually needs to do. Do not change it between the two passes.
2. Label every major criterion. Mark it as a hard constraint, core capability or search hypothesis. If the team cannot agree which category a requirement belongs to, it probably needs clarification before it becomes a filter.
3. Run the precision pass. Use the hard constraints, core capability signals and the strongest search hypotheses. Review the quality of the result set.
4. Run the discovery pass. Keep the hard constraints and core work stable. Relax selected hypotheses one at a time, widen the title family, replace exact tools with the capability they represent where appropriate, and remove industry or experience restrictions that are not genuinely fixed.
5. Compare the candidates, not just the result count. Review comparable samples and ask which search produces more credible candidates, which candidates appear only in discovery, why they were missing before, and which relaxed criterion produced useful new profiles rather than noise.
6. Keep what the market taught you. The search should become better because of evidence from real profiles, not because another requirement was automatically added to the query.

Three simple metrics for comparing the passes

Candidate count alone tells you very little. For internal search calibration, three simple working metrics can be useful. They are not universal industry benchmarks, but they help make the comparison between precision and discovery more concrete.

Qualified yield. How many candidates would you genuinely contact from the profiles reviewed? Compare the same sample size from both passes using qualified profiles divided by profiles reviewed.

Discovery lift. How many credible candidates were found only because you ran the discovery search? One practical calculation is qualified discovery-only profiles divided by all unique qualified profiles reviewed.

Boundary cost. Which search assumption excluded credible candidates? You may discover that several strong profiles disappeared only because of one exact title, one industry restriction or one preferred technology.

The important output of a discovery pass is not “we found more people”. It is “we learned which assumption was shrinking the market”.

When discovery should stop

Discovery is not a reason to keep widening a search forever. You have gone too far when most new profiles require increasingly generous interpretation to look relevant. Three warning signs are useful: the function has changed, core evidence is disappearing, or new searches add volume without adding credible candidates. Discovery should widen the route into the role, not redefine the role itself.

How to run precision and discovery in Wandify

Wandify AI Search separates semantic relevance from the criteria recruiters use to refine it. You can start from a full job description, a natural-language text query, or a reference profile through Find Similar. The AI Search uses meaning-based vector retrieval rather than relying only on exact wording, while titles, skills, keywords, experience criteria and other structured signals can be used to shape the result set.

For the product workflow and current AI Search controls, see AI Search in Wandify: A Practical Guide.

For a precision pass, start with the job description or role outcome, keep genuine Main skills, apply structural constraints that cannot change, use a focused title set where it improves relevance, and review the first candidates before tightening further.

For a discovery pass, keep the same role outcome and actual hard constraints, expand the title family where the function has multiple market labels, move substitutable frameworks or tools out of the strongest qualification signals where appropriate, and use supporting fields for context instead of turning every useful detail into a barrier.

The important distinction is that semantic search and recruiter judgment are doing different jobs. The AI can explore similarity at a scale that would be difficult to reproduce manually. The recruiter still decides which differences matter.

If you want a separate checklist for reviewing the structure of an AI-generated search, use the AI Candidate Search Audit.

Semantic search makes discovery easier, not unnecessary

There is an understandable assumption that increasingly capable AI search will eventually produce one perfect candidate list from one perfect prompt. Recruitment search does not work that neatly. Even LinkedIn’s current architecture combines multiple retrieval strategies, blending, ranking, filtering and evaluation rather than relying on one universal matching method. Its engineers describe the underlying problem as a balance between match quality and sufficient candidate coverage.

Recruiters face the same trade-off at a practical level. Make the boundaries too loose and review time increases. Make them too rigid and relevant people disappear. Semantic search improves what happens inside those boundaries. It does not remove the need to test whether the boundaries were correct.

FAQ

Does vector search make title families unnecessary?

No. Semantic search reduces dependence on exact wording, so recruiters do not need to manually predict every synonym in the way traditional Boolean search often required. Titles can still matter as structured signals or filters. A title family is useful when you deliberately want to test whether one naming convention is restricting the market. The goal is not to create the longest possible list of synonyms, but to understand whether the title boundary is helping or hurting the search.

Isn’t discovery just another name for broad search?

Not quite. A broad search simply uses fewer restrictions. A discovery search has a specific hypothesis behind it. You keep the actual role and hard constraints stable, deliberately relax one or more assumptions, then inspect the candidates that become visible. That makes discovery a calibration method rather than just a larger result page.

Does discovery mean lowering hiring standards?

No. The core capability should remain the same. What changes are the proxies used to find evidence of that capability. Accepting a Solutions Engineer who has performed the same customer-facing technical work as a Sales Engineer is not lowering the standard. Removing a truly mandatory certification would be.

Should every role have two searches?

No. Some roles have highly standardized titles, genuinely fixed qualifications and little ambiguity in how relevant experience appears. The method becomes more valuable when titles vary significantly between companies, several technologies can demonstrate the same capability, industry experience may be transferable, the role is new or changing quickly, or the first search returns a suspiciously small and uniform pool.

The best search is not the one with the cleanest result page

Semantic search has changed what recruiters can ask a candidate database to do. A recruiter no longer needs to predict every exact phrase before the search begins, because vector search can connect meaning across different terminology and surface professional experience that literal keyword matching may fail to connect. That does not make search strategy less important. It changes where the important decisions happen.

The risk is no longer only choosing the wrong keyword. It is turning an assumption into a boundary: an exact title, a familiar industry, one preferred technology, a number of years that was supposed to represent seniority, or a background that simply resembles the last person who held the role. Precision helps you find strong candidates inside the profile you expect. Discovery tests whether that profile was too narrow.

Run both when the role warrants it, then look closely at the people only the second search found. That is often where the most useful information about your search is hiding.