Private preview · MongoDB wire compatible

Enterprise vector search on commodity hardware.

Modern AI apps need vectors right alongside rich document metadata. FermionDB speaks the MongoDB wire protocol natively, delivering sub-5ms filtered vector search with a fraction of the hardware budget.

Try the playground See Rangin in action Get early access

Rangin searches 181K products with FermionDB. Try rangin.ai

Search real datasets. Inspect the exact query.

Don't just read the claims. Search The Met, arXiv, Hacker News and Wikipedia, then see the exact query and response.

Loading the playground…

Same workload. Smaller machine.

Spend your hardware on the product, not the index.

Memory

Less memory

A compact search index means the same collection fits on a smaller machine.

CPU

Less CPU

Less work per query, so the same server absorbs more traffic, especially with filters.

Recall

Same results

The finalists are ranked on their full vectors, so the savings don't cost you accuracy.

Vectors5M

Embedding dimensions

RAM price, $ per GB a month

Starts at about AWS on-demand RAM pricing. Use yours.

pgvector (HNSW): 153 GB · $879/mo
FermionDB: 83 GB · $478/mo

Estimated RAM saving

$4,811 a year · $401 a month

Estimate. Memory scales with vectors × dimensions from our memory test (199,999 vectors of 1,024 dimensions, each engine at full search speed, pgvector with an HNSW index). Your data and queries will differ.

It's its own database. It speaks MongoDB.

Use the MongoDB drivers you know with document-shaped data. Trying FermionDB means changing a connection string, not rebuilding your app.

Setup

Hands-on setup

We help each preview team get running on their own data.

Drivers

Your drivers

PyMongo, the Node driver and mongosh connect as they do to MongoDB.

Documents

Your documents

The vector is just another field, stored in the same record.

Trial

Your workload

We'll help you test it on the workload that matters to you.

Read the MongoDB API reference

met.artworks259,561 documents

Irises at Yatsuhashi (Eight Bridges)

Ogata Kōrin, 1710–1716

{
  _id: 39664,
  title: "Irises at Yatsuhashi (Eight Bridges)",
  artist: "Ogata Kōrin",
  department: "Asian Art",
  culture: "Japan",
  classification: "Paintings",
  begin_year: 1710,
  end_year: 1716,
  is_highlight: true,
  embedding: [-0.0112, 0.0008, 0.0621, … 1,152 dims]
}

Indexes

artworks_clipvector
department_1B-tree
is_highlight_1B-tree
_id_B-tree

Example: “a horse”, with the filter is_highlight: true

Search first, filter after

Grazing Horse, plate 7 from "Different Animals"

Result0 horses

FermionDB searches only the 1,751 highlights

Incense Burner of Amir Saif al-Dunya wa’l-Din ibn Muhammad al-Mawardi

Result6 horses in 7.4 ms

art = MongoClient(FERMION_URL).met.artworks
art.aggregate([{"$vectorSearch": {
    "index": "artworks_clip", "path": "embedding",
    "queryVector": clip("samurai armor with a horned helmet"),
    "filter": {"department": "Arms and Armor"},
    "numCandidates": 200, "limit": 6}}])

ok 6 results in 22.1 ms0.171  Armor (Nimai-Dō Gusoku) with Box0.167  Armor (Gusoku)0.167  Armor (Gusoku), Jo Michitaka0.165  Armor (Gusoku)0.165  Armor (Rokumai-Dō Gusoku) with Box0.163  Armor (Gusoku), Saotome Ietada

Enterprise vector search on commodity hardware.

Try FermionDB on your own data.

Tell us what you're running today and where vector search hurts. If it's a good fit, we'll get you set up during the preview.