# AI Candidate Matching

Source: https://docs.mira.day/en/docs/agent-sourcing/understanding-ai-matching

> How Mira Reasoning Embedding interprets your requirements. Goes beyond keyword search to understand context, seniority signals, and implicit criteria.



When you search for candidates on most platforms, you're doing keyword matching, type "Python developer Berlin" and get everyone with those words on their profile. It's fast, but it misses a lot.

Mira works differently. Mira Reasoning Embedding (MRE) reads your description, reasons about what you actually need, and evaluates each candidate's full career history against your requirements. Here's what that means in practice.

***

What MRE can understand that keywords can't [#what-mre-can-understand-that-keywords-cant]

"5 years of backend experience" [#5-years-of-backend-experience]

Keyword search: looks for "5 years" in the profile text.
MRE: calculates actual years from the candidate's job history timeline. If someone had 3 years at Company A and 2.5 at Company B doing backend work, MRE counts it correctly, even if their profile doesn't say "5 years" anywhere.

"Someone who's built a product from 0 to 1" [#someone-whos-built-a-product-from-0-to-1]

Keyword search: looks for the phrase "0 to 1" literally.
MRE: identifies candidates who joined early-stage companies (pre-Series A, fewer than 20 people) in product or engineering roles and stayed through a growth phase. It's reading career patterns, not keywords.

"High-growth company background" [#high-growth-company-background]

Keyword search: can't do this at all.
MRE: checks company data, funding rounds, headcount growth, revenue signals, and identifies candidates whose employers match the "high-growth" pattern.

"Public company + startup experience" [#public-company--startup-experience]

Keyword search: two separate, unrelated filters.
MRE: evaluates the full career arc, finds people who worked at a listed company AND at a startup at different points in their career, understanding that this combination signals a specific form of versatility.

What this means for you [#what-this-means-for-you]

You can describe roles the way you'd describe them to a recruiter friend:

> *"I need someone senior, technical but customer-facing, who's done enterprise sales at a startup that actually had product-market fit, not a company that just raised a lot of money."*

MRE unpacks this into multiple evaluation criteria and applies them. You don't need keyword search combinations, filters, or "search syntax."

Where the technology comes from [#where-the-technology-comes-from]

MRE comes out of Mira's own research into reasoning-based retrieval. It currently ranks first in the Embedding category and first overall on BRIGHT, a public benchmark for reasoning retrieval. Mira also built PJBenchmark, a recruiting-specific evaluation set using real hiring data.

Four papers describe the approach:

* [Time-Scaling Is What Agents Need Now](https://arxiv.org/abs/2601.02714)
* [Refine Thought](https://arxiv.org/abs/2511.13726)
* [PJB](https://arxiv.org/abs/2603.17386)
* [CRE-T1 Preview Technical Report](https://arxiv.org/abs/2603.17387)

What it doesn't do [#what-it-doesnt-do]

* **It's not magic.** If the right candidate doesn't exist in the data sources, MRE won't invent one.
* **It's not a hiring decision.** MRE surfaces the best available matches for your review. The final call is always yours.
* **Accuracy depends on data completeness.** Profiles with incomplete information produce less precise matches. Reviewing the shortlist, not just the top result, is always worth it.

Getting better results [#getting-better-results]

The quality of matching depends on the quality of your input:

* Be specific → better matches.
* Include implicit criteria → activates reasoning.
* Refine after seeing results → each round gets sharper.

See [Write an Effective JD](/en/docs/agent-sourcing/write-an-effective-jd) for practical tips.