Keyword search is less accurate on Indian resumes because it misses title and skill variants — SSE versus Software Engineer, React listed as Frontend. Semantic matching recovers those false negatives. Accuracy still needs hard must-have gates so false positives do not flood the queue. When a must-have is missing, a gate-based score is capped at 55%, even if the rest of the CV looks busy.
Keyword search fails by absence. Untamed semantic search fails by generosity. You need both recall and a gate.
Why keyword search quietly drops the right people
Naukri Boolean is a string filter. If the JD says “Software Engineer” and the resume says “SSE” or “Associate Consultant,” the profile never enters the pile. That is a false negative: a person who can do the work, invisible because of labelling. The same happens with skills: Frontend for React, Spring Boot for Java, Kubernetes buried in a project bullet as “K8s.” Recruiters who live in Resdex already know this; they keep a mental synonym list. The query does not.
Semantic matching compares meaning. SSE sits near Software Engineer. How titles get mapped is covered in how AI handles different job titles for the same role. The mechanics of Boolean versus embeddings are in keyword matching vs semantic matching in resume screening.
| Resume says | JD says | Keyword | Semantic + gates |
|---|---|---|---|
| SSE, 4 yrs | Software Engineer | Miss | Ranked, if must-haves present |
| Frontend (React in bullets) | React developer | Miss unless React is queried | Match on meaning |
| AWS, Docker, no Java | Java must-have | May hit on AWS if queried | Capped at 55% |
| Java + Spring in projects | Java, Spring Boot | Hit if strings match | Hit on related stack language |
False positives are the other half of accuracy
If you only celebrate “we found more people,” you will send clients padded profiles. Semantic overlap without a gate treats a cloud-heavy CV as close enough to a Java role. That is the equal-weight ATS mistake: counting skills instead of gating them. The structural fix — must-have versus good-to-have, and the 55% cap — is in why most ATS systems get must-have vs good-to-have wrong.
Semantic match so SSE, Consultant, and Developer can land in the same role bucket when the work matches.
A hard gate. Missing Java is not a 75% “strong” cloud profile. It is capped at 55% so the recruiter is not fooled.
Accuracy of text match is not accuracy of hire. Notice and CTC still sit on the call.
Resdex vs AI shortlisting of 64% versus 40% compares who you sourced and how you worked the channel, not a lab score for keyword versus embeddings. Manual Rex is often tighter because a recruiter already applied judgement. AI matching on a wide applicant pile is a different job. Read that comparison on its own terms in the Rex vs AI article — do not paste it onto Boolean vs semantic.
How a 55% cap changes what “accurate” means
Accuracy for an agency is: the people who should have appeared, appeared; the people who cannot do the job, did not look like 80%. The cap is the second sentence. Without it, semantic search is a generosity machine. With it, you keep the SSE you would have missed on keyword, and you do not dress up a missing must-have as a shortlist.
Take one live JD. Run your usual Naukri keyword. Then scan the same requirement with semantic rank plus must-have gates. Count how many SSEs and Consultants appeared only in the second list — and how many cloud-only CVs got capped at 55% instead of sitting at the top.
Measure accuracy the way a client measures you
Clients do not score embeddings. They shortlist or they bounce. If your AI-matched send rate sits near 45% shortlisting, the leak may be false positives (gates too soft) or confirmation (notice/CTC), not keyword syntax. If Rex still outperforms on shortlisting at 64% versus 40%, ask whether AI is ranking a noisier pool — 6,355 applications down to 325 matches is volume; volume without gates is not accuracy. Keyword search will always win on a tiny, exact-string hunt. AI matching wins when the language on Indian resumes refuses to be exact — which is most days.
Catch the SSE. Cap the false strong.
TeamOB - SafalHires combines semantic rank with must-have gates so accuracy means the right people rise — and the wrong ones cannot look like 80%.
Free trial