Prefiltering applies criteria before vector similarity search, narrowing the candidate documents and potentially improving query performance. In MongoDB Atlas Vector Search, filters are supplied through the $vectorSearch stage. section 12.3.2
Use prefiltering effectively by preparing the index, choosing supported predicates, and checking pipeline constraints.
- Index every field used for filtering by declaring it with the filter type in the vectorSearch index definition. section 12.3.2
- Filter BSON boolean, date, objectId, string, and numeric values with operators including $eq, $ne, $in, $nin, comparison operators, $and, $or, and $nor. section 12.3.2
- Respect aggregation limitations: only $and and $or are supported as aggregation-pipeline operators, and objectId fields do not support range comparisons. section 12.3.2
- Keep $vectorSearch out of view definitions, $lookup and $unionWith subpipelines, and $facet stages. section 12.3.2
Filtering a movie similarity search
A $vectorSearch query can use $or to select movies whose genre is Action or whose runtime is below 120 minutes. The queryVector searches the plot_embedding field, while numCandidates and limit control candidate and result counts. section 12.3.2
Listing 12.5 Prefiltering data with the $vectorSearch pipeline stage
db.embedded_movies.aggregate([
{
// The $vectorSearch stage performs a search based on vector similarity
"$vectorSearch": {
"index": "MongoDB-in-ActionVectorSearchIndex",
// #1
"path": "plot_embedding", // #2
"queryVector": [
-0.0016261312, -0.028070757, -0.011342932, -0.012775794,
-0.0027440966,0.008683807, -0.02575152, -0.02020668, 0.010283281,
-0.0041719596,0.021392956, 0.028657231, -0.006634482,
0.01853878, 0.0038187427,0.02959257, 0.01451522, 0.016061379,
0.000028442,0.008943722, 0.01627464, 0.024311995,
-0.025911469, 0.00022596726,-0.008863748, 0.008823762,
-0.034921836, 0.007910728,0.01515501,0.019300312,0.010749794,
0.0075308536,0.018287312,0.016567878,0.012869096,0.015528221,
0.0078107617,0.003415542,0.025911469,0.00022596726,
0.0088637480.011156326,0.013522214,0.020646530.0121601,
0.011596181,0.017247654,0.0005939711,-0.026977783,-0.003942035,
0.007890735,0.009690142,0.02020668,0.0012962399,0.025911469,
0.00022596726,0.0088637480.003415542,0.00024970944,0.012775794,
0.0114762215,0.013422247,0.010429899,0.03686786,0.006717788,
0.027484283,0.011556195,0.036068123,0.013915418,0.02020668,
0.001292399,0.02020668,0.001292399,0.003415542,0.003415542,
0.0016327957,0.0151016945,0.020473259,0.004671795,0.012555866,
0.0209531,0.01982014,0.024485271,0.0105431955,0.005178295,
0.033162415,0.013795458,0.007150979,0.010243294,0.005644808,
0.0045618312,0.0024725192,0.004305249,0.008197301,0.0014203656,
0.0018460588,0.005015015, 0.02020668,-0.0012962399,-0.003415542
], // The shortened query embedding
"filter": {
"$or": [
{ genres: "Action" },// #3
{ runtime: { "$lt": 120 } } // #4
]
},
"numCandidates": 150,// #5
"exact": false, // #6
"limit": 5 // #7
}
},
{
// The $project stage specifies which fields to include in the output
"$project": {
"_id": 0, // #8
"plot": 1, // #9
"title": 1, // #10
"score": { $meta: "vectorSearchScore" } // #11
}
}
])
Inspect the aggregation example to see where the filter, vector field path, query vector, and search controls belong in $vectorSearch. section 12.3.2
from pymongo import MongoClient
def vector_search():
# MongoDB connection string
uri = 'your_mongodb_connection_string'
# Create a new MongoClient
client = MongoClient(uri)
try:
# Connect to the MongoDB server
client = MongoClient(uri)
# Select the database and collection
database = client['your_database_name']
collection = database['your_collection_name']
# Define the pre-filters and vector search query
query = {
"$and": [
{"category": "electronics"}, # Example pre-filter
{"price": {"$lt": 500}}, # Another example pre-filter
{
"$vectorSearch": {
"queryVector": [/* your query vector here */],
"numResults": 10,
"path": "your_vector_field_name"
}
}
]
}
try:
# Execute the vector search query
➥ # and convert the results to a list
results = list(collection.find(query))
# Print the results
for result in results:
print(result)
except Exception as query_error:
# Handle errors that occur during the query execution
print('Error executing vector search query:', query_error)
except Exception as connection_error:
# Handle errors that occur while connecting to MongoDB
print('Error connecting to MongoDB:', connection_error)
finally:
try:
# Ensure the MongoDB client is closed
client.close()
except Exception as close_error:
# Handle errors that occur while closing the MongoDB connection
print('Error closing MongoDB connection:', close_error)
# Call the vector_search function
vector_search()
The Python pattern combines category and price prefilters with $vectorSearch, executes the query through pymongo, prints results, and includes connection error handling. section 12.4.2
