Your best candidate probably never typed the words in your job description. Metix AI searches by meaning instead of by string, so what comes back is people who do the work, whatever they happened to call it.
Keyword search finds people who used your words. Semantic sourcing finds people who do the work. Mira, the Metix AI recruiter, searches 860M+ profiles in 190+ countries with an embedding model we trained for hiring, and on a pool of 300 real job descriptions that model lifted Recall@50 over the general-purpose baseline, from 68.89% to 77.55%. One honest caveat travels with the claim: we expect the gains to concentrate on long-tail roles, where words and work diverge. A mainstream role with thousands of obvious matches should gain less, because keyword search was never the bottleneck there.
Most sourcing tools treat the search box as a string-matching problem. You write Boolean lines, the engine scans millions of profiles for those tokens, and back comes everyone who typed the same words you did. That sounds neutral. It is anything but. It quietly turns your search into a vocabulary test, and the people you most want to meet never agreed to sit it.
We catalogued the failures while building Mira, and they repeat with numbing regularity. A general-purpose engine matches SOC 2 compliance experience to SoC hardware design, since the string agrees even though the work could not be further apart. An enterprise product marketing role pulls in brand marketing resumes, because the two share a generic vocabulary. And a senior distributed systems search ranks junior profiles high simply because the keyword Go appears. In every case the token is present and the fit is absent. Our system report on Mira walks through each of these.
The human side makes it worse. Titles inflate. The same title carries different scopes in different industries. Strong people describe their careers in project language, not recruiter language, and the long tail of skill synonyms and abbreviations never ends. A bigger pool amplifies a vocabulary problem; it does not fix it.
An embedding model reads a piece of text and places it as a point in a space with thousands of dimensions, arranged so that distance tracks meaning. Two texts land near each other when they describe the same thing, even if they share no words at all. A profile that says ran on-call for a payments platform sits close to a job description asking for a site reliability engineer, because the model has learned that those sentences describe the same job. Search stops being a string hunt. It becomes a question of which profiles sit nearest to the work you described.
Off-the-shelf embeddings get you part of the way and then stall, because they measure topical resemblance rather than hiring relevance. So we trained our own. Mira-Embeddings-V1, our recruitment reranking model, starts from real job descriptions and uses a five-stage prompt pipeline to synthesize diverse positives and hard negatives, then applies two rounds of LoRA adaptation: contrastive training between job descriptions first, then triplet alignment between job descriptions and resumes. A final BoundaryHead reranker separates the cases that sink keyword tools most reliably, roles that share a title but differ in scope, seniority, or function. All of that supervision was synthesized by LLMs from a modest set of real job descriptions, without large-scale manual labeling.
Trained this way, similarity starts to encode what hiring actually turns on: role-specific terminology, the fine-grained differences between adjacent roles, and relevance calibrated by seniority and hard constraints. A junior profile with the right keyword is no longer scored on the keyword alone; the training pushes actual fit above token presence.
For the person reading the slate, the effect is simple. Candidates are ranked by how close their work sits to your role, however they chose to describe it.
Sourcing is best judged on recall under a review budget. Nobody reads ten thousand profiles. You read the top fifty, so the question that matters is how many of the genuinely qualified people appear inside that window. The shape of the system matters for reading the results: in production, an upstream retriever pulls a candidate shortlist for each role, and the embedding model reranks that shortlist so qualified people rise into the window a human will review.
On a pool of 300 real job descriptions with candidates returned by that production retriever, mira-embeddings-v1 raised Recall@50 from 68.89% to 77.55%, and Precision@10 from 35.77% to 39.62%. On a larger, LLM-judged supportive protocol over 44,138 candidates, Recall@200 rose from 0.5969 to 0.7047. Same pool, same roles, different ranking. The qualified people were always in there. More of them now surface where someone will look.
Retrieval metrics are the honest baseline, and the business numbers follow from them. Against a general-purpose embedding baseline, the fine-tuned model produced an improvement of over 40% on our primary business metric, and more than 90% savings in operational cost per qualified, interested candidate in production-like evaluation. Fewer wrong profiles retrieved means fewer profiles to push through review and outreach for every interview that happens. Ranking quality sets the cost structure of the whole pipeline.
Here is the part a feature page is supposed to skip. Nothing in the mechanism says a match on meaning should beat a match on words by the same margin everywhere, and pretending otherwise would sell it short. Think about what each approach needs to succeed. Keyword search works exactly when vocabulary is a good proxy for the work. On a mainstream role, a staff accountant, a React developer in a major city, the market has already standardized the words. Thousands of people hold the title, describe it the same way, and match on both string and substance. Keyword search surfaces plenty of them. We expect semantic ranking to order that list better, but the marginal gain should be modest, because finding people was never the hard part of that search.
Long-tail roles are the opposite case: a compensation lead for a company expanding into three new countries, say, or a firmware engineer who has shipped under medical-device audit. For roles like these, the title is nonstandard or shared with unrelated jobs, the skills hide inside project descriptions, and the people you want may describe their work in another industry’s dialect. Vocabulary stops being a proxy for anything. This is where a match on meaning changes the slate you see, and where a match on words quietly returns the wrong crowd with confidence.
Our published numbers are averages across pools of real job descriptions; we have not published a per-role-type breakdown, so we will not quote one. But the mechanism tells you where to expect the gains to concentrate: wherever the words and the work diverge. If everything you hire is short-tail, you would pick Metix AI for the outreach and scheduling rather than for the search. If you hire anything rare, the sourcing gap is the product.
The matching runs over one of the largest talent pools assembled for recruiting: more than 860 million profiles across 190+ countries, continuously refreshed. Every record is built exclusively from publicly available sources, the professional presence people choose to publish, and the data pipeline is built to comply with GDPR and CCPA. No private APIs, no restricted content.
Scale and matching compound here. A rare combination of skills is thin on the ground in any one market, by definition. Spread the search across 190+ countries and the set of people who genuinely fit becomes large enough to be worth ranking well, which is exactly the situation semantic ranking is best at.
The search itself is a loop. Hiring intent is almost always underspecified at the start, so Mira’s Search Agent treats each round of retrieval as evidence. A pool that comes back too small gets broadened in a controlled way; one that comes back too large or too weak gets tightened. Hard constraints and soft preferences never blur into each other, and every revision is logged. When a request is internally contradictory, the agent surfaces the conflict instead of broadening forever. The point of all that machinery is mundane and valuable: the search converges on what you meant, not on what you first typed.
A great ranked list is still a list. On its own it hands you the outreach, the screening, and the scheduling, which is exactly the trap we wrote about in AI sourcing tool vs AI recruiter. Inside Metix AI, the sourcing stage feeds the rest of Mira’s end-to-end pipeline: a Match Agent grades each candidate against your actual requirements with evidence attached, an Engage Agent runs the outreach you approve, and what reaches you is interested candidates with interview slots. Across our early enterprise customers, 95% of roles have the first interview scheduled within 24 hours.
Retrieval quality also does not get to drift silently. Changes to Mira’s models and prompts run against evaluation suites before they ship, and a key metric falling past its threshold blocks the release, a discipline we lay out in Agent Evaluation, Done Right. The numbers on this page are the kind that gate our own merges.
The pricing follows the same logic. You never pay for profiles or contact reveals. A credit is spent only on an interested, hire-ready candidate, on plans from $49 to $299 a month. Better retrieval makes that meter cheaper to run, which is why we publish the retrieval numbers at all. Search quality is what makes the unit economics of paying per candidate work.
Semantic sourcing is easiest to judge on a role that has already beaten a keyword tool. Bring that one. You write the brief, Mira runs the search across 860M+ profiles, and you judge the slate that comes back. If nobody clears your bar, no credit is spent.
Semantic sourcing matches candidates to roles by meaning rather than by keywords. Mira, the Metix AI recruiter, searches 860M+ profiles in 190+ countries with mira-embeddings-v1, an embedding model trained on hiring data, so people who describe the same work in different words still surface.
On a pool of 300 real job descriptions with candidates returned by the production retriever, mira-embeddings-v1 lifted Recall@50 over the general-purpose baseline from 68.89% to 77.55%, and Precision@10 from 35.77% to 39.62%. On a larger, LLM-judged supportive protocol over 44,138 candidates, Recall@200 rose from 0.5969 to 0.7047.
We expect it to help less than on rare ones. Metix AI has not published a per-role-type breakdown, but the mechanism predicts the gains concentrate on long-tail roles, where titles and vocabulary diverge from the actual work. A mainstream role with thousands of obvious matches should gain only modestly, because keyword search was never the bottleneck there.
Every record is built exclusively from publicly available sources across 190+ countries, and the data pipeline is built to comply with GDPR and CCPA. Metix AI uses no private APIs and no restricted content.
No. You never pay for profiles or contact reveals. A credit is spent only on an interested, hire-ready candidate, on plans from $49 to $299 a month, and the free 14-day trial includes 3 New Roles and 12 Credits with no card required.