Docs · API reference

MongoDB API reference.

FermionDB speaks the MongoDB wire protocol, so you use the MongoDB drivers and query language you already know. This page covers connecting, indexing and vector search, and lists what FermionDB supports.

Connect

Point any MongoDB driver at your FermionDB address. Nothing else changes.

from pymongo import MongoClient

client = MongoClient("mongodb://HOST:PORT")
products = client.shop.products

Create the indexes

One vector index per collection, plus a regular index on each field you filter by. Create them before you load documents.

# Regular indexes on the fields you filter by
products.create_index([("in_stock", 1)])
products.create_index([("category", 1)])

# The vector index
client.shop.command({
    "createSearchIndex": "products",
    "name": "products_vec",
    "type": "vectorSearch",
    "definition": {"fields": [{
        "type": "vector", "path": "embedding",
        "numDimensions": 1536, "similarity": "cosine",
    }]},
})

$vectorSearch

The first stage of an aggregation pipeline. Later stages work on its results as usual.

results = products.aggregate([
    {"$vectorSearch": {
        "index": "products_vec",
        "path": "embedding",
        "queryVector": embedding,          # 1,536 numbers, same as the index
        "numCandidates": 200,
        "limit": 10,
        "filter": {"in_stock": True, "category": {"$in": ["shoes", "bags"]}},
    }},
    {"$project": {"name": 1, "price": 1, "score": {"$meta": "vectorSearchScore"}}},
])

$vectorSearch parameters
Parameter	Type	Meaning
index	string	Name of the collection's vector index.
path	string	Field that holds the vector.
queryVector	number[]	The query embedding. Same length as the index's numDimensions.
numCandidates	integer	How many candidates the search considers before returning the best. Higher is more thorough.
limit	integer	How many documents to return.
filter	document	Optional. Conditions on indexed fields, checked during the search.

Each result's similarity score is available as { $meta: "vectorSearchScore" } in a $project stage.

Supported MongoDB API

What FermionDB supports today. Need something that isn't listed? Tell us when you request early access.

Drivers: Any MongoDB driverPyMongoNode.jsGoJavamongosh
Pipeline stages: $vectorSearch$match$project$group$sort$limit
$vectorSearch filter: Equality$in$nin$ne
$match operators: Equality$in$gte$regex
Expressions: $sum$meta: "vectorSearchScore"
Commands: insertManyaggregatecreateIndexcreateSearchIndexlistSearchIndexes
Indexes: Vector (cosine)Single-field

Try it in the playground Get early access