Overview Playground Games Pricing Dashboard Sign in Create account
info@mirrorfractal.com
Beta · Mirror Fractal Lab · Patent Pending MMXXVI · v1.0.0

Close enough is exactly right.

The Fuzzy is an approximate retrieval engine. Typos, partial hashes, noisy sensor readings, mutated DNA — it finds the neighbourhood anyway. Built on the Mirror Trie Binary, a patent-pending data structure that turns approximate search into a bounded tree traversal.

Scroll to explore
Fuzzy search Binary radix Mirror trie Hamming distance Multi-probe DNA retrieval Unicode ready Sub-150ms Patent pending

Measured, not promised.

Rust core, release build with LTO, single thread, arena allocator, 500 queries per data point. BK-Tree is the baseline.

@ 10M keys
14.9×
faster than BK-Tree · multi-probe
Recall
99.8%
multi-probe at distance d ≤ 2
Latency
138ms
on 10M keys, single thread
Memory
+33%
overhead vs flat table

Three passes.
Bounded by the bits.

A Mirror Trie Binary stores hashes as paths, one bit per level. Each node is a mirror — branching left and right on the same decision. Approximate search becomes a guided descent with an error budget.

  1. Phase 1
    Exact descent.

    Walk the trie bit by bit. If the target is present, it lands in O(hash-length). If not — exploration begins from the last matching node.

  2. Phase 2
    Mirror exploration.

    At each node along the path, allow up to d − 1 further flips. The error budget shrinks with depth; the tree prunes itself.

  3. Phase 3
    Multi-probe union.

    Flip each high-bit independently, union the results. This is where the 99.8% recall comes from — and the cost is still sub-linear in the dataset.

Three ways to see it run.

Open playground →

Anywhere perfect input is a lie.

Security
Leaked-password checks.

Match a password against billions of breached hashes — with typo tolerance. One bit flipped ≠ secure.

Bio
DNA 1-SNP retrieval.

Radix-4 variant. Mathematically proven 100% recall for single nucleotide polymorphism search. 82× speedup at 500K.

Search
Autocomplete with typos.

Users mis-type. Your search layer doesn't have to punish them. Pull candidates within edit distance 2.

Forensics
Fuzzy file hashes.

Find mutated malware samples, re-compressed media, near-duplicate documents — across TrueFile / PhotoDNA / ssdeep spaces.

Free to start.
Pay when it scales.

See all plans →
Free
$0/mo

For weekend experiments.

  • 10 K queries / month
  • 1 index, up to 100 K keys
  • Community support
Enterprise
Talk

Air-gapped, on-prem, custom.

  • On-premise binaries
  • Patent licensing included
  • Dedicated engineering
"Every query is a guess.
The trick is to bound the guessing."