Automated Embedding Overview

Enable and Use Automated Embedding

1
{
2
  "fields": [
3
    {
4
      "type": "autoEmbed",
5
      "modality": "text",
6
      "path": "<field-to-index>",
7
      "model": "<embedding-model>"
8
    },
9
    {
10
      "type": "filter",
11
      "path": "<field-to-index>"
12
    },
13
    ...
14
  ]
15
}

1
[
2
  {
3
    "$vectorSearch": {
4
      "index": "<index-name>",
5
      "path": "<field-to-index>",
6
      "query": "<query-text>",
7
      "model": "<embedding-model>"
8
    }
9
  },
10
  {
11
    "$project": {
12
      "_id": 0,
13
      "<field-to-index>": 1,
14
      "score": { "$meta": "vectorSearchScore" }
15
    }
16
  }
17
]

Embeddings Storage

Automated embedding indexes generate vector embeddings asynchronously and persist to your MongoDB cluster on a separate reserve database. Each Automated Embedding index has exactly one corresponding generated embeddings collection. The generated embeddings collection is stored in a dedicated internal database on the same cluster.

To learn more, see Generated Embeddings Collection.

Available Models

MongoDB Vector Search integrates with Voyage AI's state-of-the-art embedding models, each optimized for specific use-case:

To learn more, see Models for Automated Embedding.

Key Concepts

embedding model

Embedding models are algorithms that convert data into vector embeddings that capture your data's semantic, or underlying, meaning. These vectors enable vector search.

To learn more about the embedding models for Automated Embedding, see Available Models.

vector embeddings

A vector embedding is an array of numbers, with each dimension representing a different feature or attribute of your data. Vectors can be used to represent any type of data, from text, images, and video to unstructured data. You create vector embeddings by passing your data through an embedding model, and you can store these embeddings in a database that supports vector embeddings like MongoDB.

To learn more about the embeddings storage for Automated Embedding, see Generated Embeddings Collection.

tokens

In the context of embedding models and LLMs, tokens are the fundamental units of text, such as words, subwords, or characters, that the model processes to create embeddings or generate text. Tokens are how you are billed for usage of embedding models and LLMs.

To learn more about tokens for Automated Embedding, see Manage Billing for Automated Embedding.

rate limits

Rate limits are restrictions imposed by API providers on the number of requests a user can make within a specific time frame, often measured in tokens per minute (TPM) or requests per minute (RPM). These limits ensure fair usage, prevent abuse, and maintain the stability and performance of the service for all users.

To learn more about rate limits for Automated Embedding, see Rate Limits.

quantization

Quantization reduces the precision of vector embeddings to decrease memory and storage usage, with trade-offs in search accuracy. For Automated Embedding, MongoDB Vector Search supports the following quantization types:

Quantization Type	Description
float	Stores the vector embeddings as 32-bit float values.
scalar	Reduces each vector dimension from 32-bit float to 8-bit integer.
binary	Reduces each vector dimension to a single bit and rescores the top results.
binaryNoRescore	Reduces each vector dimension to a single bit without rescoring.

To learn more about the quantization for Automated Embedding, see About Quantization.

number of dimensions

The number of dimensions specifies the length of the embedding vector for each document (how many numbers are in the array). Higher dimensions capture more semantic detail and generally improve retrieval accuracy, but increase storage and compute costs (index size, RAM usage, and sometimes latency).

To learn more, see numCandidates Selection.

similarity

The similarity function is used to measure the similarity between two vectors or the closeness of a query vector to the vectors in the index. MongoDB Vector Search supports the following similarity functions:

• cosine - measures similarity based on the angle between vectors.

• dotProduct - measures similarity like cosine, but takes into account the magnitude of the vector. We recommend this similarity function for full-fidelity or scalar quantization.

• euclidean - measures the distance between ends of vectors. We recommend this similarity function for binary or binaryNoRescore quantization, where vectors are compressed and distance in Hamming or Euclidean space is the right signal.

To learn more about the similarity functions for Automated Embedding, see About the Similarity Functions.