Natural-language candidate sourcing is a way to find candidates by describing the person you want in plain language, instead of building a Boolean string or setting filters by hand. The recruiter types a sentence, the system works out the criteria behind it, and it retrieves and ranks candidates by meaning rather than exact keyword matches. Many sourcing tools now offer some version of it. This guide explains how it works and how to tell a strong implementation from a weak one.
What natural-language candidate sourcing is
Instead of writing ("software engineer" OR "developer") AND fintech AND ("Series B" OR "Series C"), the recruiter types "senior backend engineer in San Francisco who worked at a Series B fintech and knows AWS and Golang" and the system does the translation. See our guide to natural language searches for more examples.
The defining property is retrieval by meaning. The system reads the intent behind the words, recognizes that "lead dev," "engineering manager" and "senior software engineer" can describe similar seniority, and returns people whose experience fits even when their job titles use different words. Boolean search matches the characters you typed. Natural-language search matches the concept you meant.
How it works, step by step
A natural-language sourcing query goes through four stages. The stages describe the general pattern, and the examples show how Juicebox implements it.
Plain-language prompt. The recruiter describes the role and the ideal candidate in a sentence, the way they would explain it to a colleague. There are no operators and no field-by-field form.
Inferred criteria. The system parses the sentence into structured requirements such as role, seniority, industry, location, skills and company stage. Two AI techniques do the work. Semantic search compares the meaning of the query with the meaning of each profile, and entity recognition breaks the request into structured fields.
Retrieval by meaning. The system searches a candidate index and returns people whose experience matches the inferred criteria, including candidates whose titles differ from the query but whose background fits. Juicebox searches 800M+ profiles from 30+ data sources.
Ranked results. Candidates come back ranked by fit, with the reasoning shown, rather than as an unsorted keyword list. Juicebox writes a summary on each profile explaining why it matches, and it evaluates up to 5,000 profiles against your criteria to find the best fits.
How it differs from Boolean search and filter-based databases
Natural-language sourcing, Boolean search and filter-based candidate databases all answer the same question, "who should I reach out to?", in different ways.
Input: A plain-language sentence describing the person
What it matches: Meaning and intent, including related terms and titles you did not type
Who can use it well: Any recruiter who can describe the role
Speed to a usable list: Fast: a ranked list from the first sentence
Control and precision: High, when criteria are editable and you can add filters
Main risk: Results that are too broad if criteria cannot be tightened
Best for: Complex or loosely defined roles, career changers, recruiters without deep Boolean skills The three are not mutually exclusive. Most natural-language tools let a recruiter start with a sentence and then add filters for precision. The sentence does the broad, intent-aware retrieval, and the filters narrow it to the hard constraints.
How to evaluate a natural-language sourcing tool
Accepting a sentence as input does not make a tool good. Two products that both take plain language can differ sharply in what they find and how well they rank it. These five questions separate them.
How big and fresh is the index? How many profiles does it search, from how many sources, and how current are they? Coverage sets the ceiling on what any query can find. Juicebox searches 800M+ profiles from 30+ data sources.
Does retrieval find the right people? Does it surface strong candidates whose titles do not match the query? Published benchmarks help. On Exa's People Search Benchmark of 1,400 real queries, Juicebox scored 79%, according to our search announcement.
Can you see why each candidate ranked where they did? Visible reasoning lets a recruiter trust and correct the order instead of re-sorting by hand. Juicebox shows the criteria it used and a written explanation for each candidate.
Can you edit the criteria? When the system infers the wrong thing, can you correct it and re-rank, or are you stuck with its first reading? Editable criteria give a natural-language tool the precision of a filtered search.
Where do the results go? Does the tool stop at a list, or carry candidates into review, outreach and your ATS? Sourcing is one step in a hiring workflow, and the handoff matters.
Where filters and depth still matter
Natural-language sourcing does not replace every kind of search, and a tool that claims it does is the wrong tool. Structured methods still work better in three cases.
Hard, binary requirements. When a role demands an exact attribute, such as a security clearance, a licensed certification or a location radius, an explicit filter is the most reliable way to enforce it. Retrieve broadly with a sentence, then apply filters for the must-haves.
Deep technical signals. Some roles depend on signals that live outside a standard profile, such as open-source contributions, publications or patents. Platforms differ in how much of this they index. Juicebox includes activity signals such as GitHub contributions and publications, and you can filter by developer data.
Precise keyword searches. When a recruiter already knows the exact terms that define a fit, a literal Boolean string can be faster and easier to control than describing the same thing in prose.
Natural-language sourcing earns its place on searches where intent is loosely defined, titles vary and speed matters, and it pairs with filters for the constraints that must be exact.
How Juicebox does it
In Juicebox Search, a recruiter describes the role in plain language, and Juicebox configures the filters and runs the search across 800M+ profiles from 30+ data sources. It works out the quality criteria from the prompt, ranks matches in the same step, and shows the reasoning behind each rank. The same search powers Juicebox Agents, which run it around the clock, learn from the candidates you approve or reject, and either email candidates or add them to a shortlist for review.
CoinTracker, a Juicebox customer, cut sourcing time by 50% and saves 30+ hours a week with Juicebox Agents. Read the CoinTracker story.
Common questions
What is natural-language candidate search?
It lets a recruiter find candidates by typing a plain-language description of who they want, instead of building a Boolean string. The system reads the intent behind the sentence, works out the criteria, and retrieves and ranks candidates by meaning, including people whose job titles differ from the query but whose experience fits.
How is natural-language search different from Boolean search?
Boolean search matches the exact keywords and operators a recruiter types, so it can miss candidates who describe the same experience in different words. Natural-language search matches meaning and intent, so it surfaces those candidates, and the recruiter does not need to know Boolean syntax.
Is natural-language search better than Boolean?
Each fits different searches. Natural-language search is faster and finds more relevant candidates for complex or loosely defined roles, and for recruiters without deep Boolean skills. Boolean search gives more control when a recruiter already knows the exact must-have keywords. Many tools let you start with a sentence and add filters to get both.
Does natural-language search replace filters?
No. Filters are still the most reliable way to enforce hard requirements such as a location radius, a certification or a security clearance. The strongest pattern is to retrieve broadly with a prompt and then apply filters for the must-haves.
What should I look for in a natural-language sourcing tool?
Look at index size and freshness, retrieval quality (ideally shown by a benchmark), ranking with visible reasoning, editable criteria, and where the results go next: review, outreach and your ATS.
