# Vector Search

Typesense has the ability to index embeddings generated by any machine learning model, and then do a nearest-neighbor (KNN) search on this data.

# Use-cases

Here are some example use-cases you can build, using vector search as the foundation:

  1. Semantic search
  2. Recommendations
  3. Hybrid search (Keyword Search + Semantic Search + Filtering)
  4. Visual image search
  5. Integrate with LLMs (opens new window), to get them to respond to queries using your own dataset (RAG)

You can also combine any of the above with features like filtering, faceting, sorting, grouping, etc to build a user-friendly search experience.

# What is an embedding?

An embedding for a JSON document is just an array of floating point numbers (eg: [0.4422, 0.49292, 0.1245, ...]), that is an alternate numeric representation of the document.

These embeddings are generated by Machine Learning models in such a way that documents that are "similar" to each other (for different definitions of similarity depending on the model used), have embeddings that are "closer" to each other (cosine similarity).

Here are some common models you can use to generate these document embeddings:

  • Sentence-BERT
  • E-5
  • CLIP
  • OpenAI's Text Embeddings model
  • Google's PaLM API
  • Google's Vertex API

You can import embeddings generated by these models into Typesense into a special vector field and then do a nearest neighbor search, giving another set of vectors or a document ID as the input, and get the documents that are closest (cosine similarity) to your input.

You can also have Typesense generate these embeddings for you, using OpenAI, PaLM API or one of the built-in ML models listed here (opens new window).

# Live Demo

Here is one (of many possible) practical applications of vector search - a "Find Similar" feature in an ecommerce store: ecommerce-store.typesense.org (opens new window). (Click on Find Similar below each product).

# Read More

Here are two articles that talk about embeddings in more detail:

Let's now discuss how to do index and search embeddings in Typesense.

# Index Embeddings

# Option A: Importing externally-generated embeddings into Typesense

If you have already generated embeddings using your own models outside Typesense, you can import them into Typesense.


Here's (opens new window) a quick example of how to use the Sentence-BERT model to generate embeddings outside Typesense.

Once your document embeddings are ready, you want to create a collection that contains a float[] field with a num_dim property for indexing them. The num_dim property specifies the number of dimensions (length of the float array) that your embeddings contain.

Let's create a collection called docs with a vector field called embedding that contains just 4 dimensions.


We're creating a vector with 4 dimensions in the examples to keep the code snippets readable.

Depending on what model you use, real world use will require creating vector fields with at least 256 dimensions to produce good results.

Let's now index a document with a vector.

# Option B: Auto-embedding generation within Typesense

To simplify the process of embedding generation, Typesense can automatically use your JSON data and either OpenAI API, PaLM API or any of the built-in embedding models listed here (opens new window) to generate & store embeddings.

When you do a search query on this automatically-generated vector field, your search query will be vectorized using the same model used for the field, which then allows you to do semantic search or combine keyword and semantic search to do hybrid search.

# Creating an auto-embedding field

To create a field that automatically embeds other string or string array fields, you need to set the embed property of the field.

Here's an example:

In this example the embedding vector field will be generated automatically while indexing a document, using the concatenated values of the product_name and categories fields (separated by spaces).

# Using Built-in Models

These models are officially supported by Typesense and stored in the Typesense Hugging Face repository here (opens new window).

You can specify them by adding the ts namespace before the model name. Typesense will automatically download these models and make them available for use when you index documents after creating the collection.

When you create a collection with the schema above, the all-MiniLM-L12-v2 model will be downloaded and your documents will be automatically embedded by this model and will be stored in the embedding field.

See our Hugging Face repo (opens new window) for all officially supported models. If you need support for additional publicly-available models, feel free to convert the model to ONNX format and send a PR to our Hugging Face models repo (opens new window).

# Using a GPU (optional)

Embedding models are computationally intensive to run. So when using one of the built-in models, you might want to consider running Typesense on a server with a GPU to improve the performance of embedding generation, especially for large datasets.

# On Typesense Cloud:

For select RAM / CPU configurations (opens new window), you'll find the option to turn on "GPU Acceleration" when provisioning a new cluster or under Cluster Configuration > Modify for Typesense versions 0.25.0 and above.

# When Self Hosting:

You would have to install the following additional dependencies, after which Typesense will automatically make use of any available Nvidia GPUs:

  1. Install CUDA following the instructions on Nvidia's site here (opens new window).

    You want to specifically install the following versions of these packages: cuda=11.8.0-1, libcudnn8= and libcudnn8-dev= and their dependencies.

  2. Install cuDNN following the instructions here (opens new window).

    You want to specifically install the libcudnn8 and libcudnn8-dev packages.

  3. Add the following to /etc/profile.d/cuda-path.sh:

    export PATH=/usr/local/cuda/bin${PATH:+:${PATH}}
    export LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:/usr/local/cuda/lib64
    export CUDA_HOME=/usr/local/cuda
  4. Install the Typesense GPU dependencies package:

# Using OpenAI API

You can also have Typesense send specific fields in your JSON data to OpenAI's API to generate text embeddings.

You can use any of OpenAI models listed here (opens new window).

When you create the collection above, we will call the OpenAI API to create embeddings from the product_name field and store them in the embedding field every time you index a document.

You have to provide a valid OpenAI API key in model_config to use this feature.

# Using Google PaLM API

This API provided by Google MakerSuite (opens new window) to generate embeddings.

Note: The only supported model is embedding-gecko-001 for now.

# Using GCP Vertex AI API

This API also provided by Google under the Google Cloud Platform (GCP) umbrella.

You would need the following authentication information to use this method:

  • GCP access token (must be valid while creating the field)
  • GCP refresh token
  • GCP application client ID
  • GCP application client secret
  • GCP project ID

Please refer to the Vertex AI docs for more information on how to fetch these values.

# Remote Embedding API parameters

You can use any of the following parameters to fine-tune how API calls are made to remote embedding services:

Parameter Description Default
remote_embedding_timeout_ms How long to wait until an API call to a remote embedding service is considered a timeout, during a search 30s
remote_embedding_num_tries The number of times to retry an API call to a remote embedding service on failure, during a search 2

# During Indexing

Parameter Description Default
remote_embedding_batch_size Max size of each batch that will be sent to remote APIs while importing multiple documents at once. Using lower amount will lower timeout risk, but increase number of requests made. 200
remote_embedding_timeout_ms How long to wait until an API call to a remote embedding service is considered a timeout during indeixng. 60000
remote_embedding_num_tries The number of times to retry an API call to a remote embedding service on failure during indexing. 2

# Using your own models

You can also use your own models to generate embeddings from within Typesense. They must be in the ONNX file format.

Create a directory under <data_dir>/models and store your ONNX model file, vocab file, and a JSON for model config there.

Note: Your model file MUST be named as model.onnx and the config file MUST be named as config.json.

# Model config file

This file will contain information about the type of model you want to use.

The JSON file must contain model_type (type of the model; we support bert and xlm_roberta at the moment) and vocab_file_name keys.

Directory Structure:


Contents of config.json:

    "model_type": "bert",
    "vocab_file_name": "vocab.txt"

Create an embedding field using the directory name as model_name in model_config.

# Optional Model Parameters

These are optional model parameters, which may be required to use with your custom models.

# Indexing prefix and query prefix

Some models may require a prefix to know if texts are queries or they are actual texts to query on (you can check intfloat/e5-small, for example).

If you set this property in model_config, the given indexing prefix will be added to the text that will be used to create embeddings when you index a document and query_prefix to the actual query before creating embeddings of it.Example:

For this example, when you index a document:

   "product_name": "ABCD"

The text used to generate embeddings for the embedding field will be passage: ABCD instead of ABCD. And when you query, if your query is EFGH, it will be embedded as query: EFGH instead of EFGH.

Once you've indexed your embeddings in a vector field, you can now search for documents that are "closest" to a given query vector.

To control the number of documents that are returned, you can either use the per_page pagination parameter or the k parameter within the vector query.

NOTE: If both per_page and k parameters are provided, the larger value is used.


Since vector search queries tend to be large because of the large dimension of the query vector, we are using the multi_search end-point that sends the search parameters as a POST request body.

Every matching hit in the response will contain a vector_distance field that indicates how "close" the document's vector value is to the query vector. Typesense uses the cosine similarity, so this distance will be a value between 0 and 2.

  • If the document's vector perfectly matches the query vector, the distance will be 0
  • If the document's vector is extremely different from the query vector, then the distance will be 2.

The hits are automatically sorted in ascending order of the vector_distance, i.e. best matching documents appear first.

Sample Response

# Querying for similar documents

If you have a particular document id and want to find documents that are "similar" to this document, you can do a vector query that references this id directly.

curl 'http://localhost:8108/multi_search' \
  -X POST \
  -d '{
        "searches": [
            "collection": "docs",
            "q": "*",
            "vector_query": "embedding:([], id: foobar)"
# Be sure to replace `embedding` with the name of the field that stores your embeddings.

By specifying an empty query vector [] and passing an id parameter, this query would return all documents whose embedding value is closest to the foobar document's embedding value.


The foobar document itself will not be returned in the results.

When using auto-embedding, you can directly set query_by to the auto-embedding field to do a semantic search on this field.

Typesense will use the same embedding model that was used to generate the auto-embedding field to generate vectors for the q parameter and then do a nearest neighbor search internally.

This will automatically embed the chair query with the same model used for the embedding field and will perform a nearest neighbor vector search.

When using auto-embedding, you can set query_by to a list of both regular fields and auto-embedding fields, to do a hybrid search on multiple fields.

Typesense will do a keyword search on all the regular fields, and a semantic search on the auto-embedding field and combine the results into a ranked set of results using Rank Fusion:

K = rank of document in keyword search
S = rank of document in semantic search

rank_fusion_score = 0.7 * K + 0.3 * S

The 0.7 and 0.3 values can be changed using the alpha parameter.

During hybrid search, the _text_match clause in sort_by will refer to the combined fusion score.

You can also do a hybrid search when using your own embedding field, by combining the q parameter with the vector_query parameter.

Typesense will do a keyword search using the q parameter, and a nearest neighbor search using the vector_query field and combine the results into a ranked set of results using Rank Fusion (see above).

# Weightage for Semantic vs Keyword matches

By default, Typesense assigns a weight of 0.3 for vector search rank and a weight of 0.7 for keyword search rank. You can adjust the weight assigned to vector search ranking via the alpha option of the vector_query parameter.

For example, to set a weight of 0.8 to vector search ranking, set alpha to 0.8:

curl 'http://localhost:8108/multi_search' \
    -X POST \
    -d '{
          "searches": [
              "collection": "products",
              "query_by": "embedding,product_name",
              "q": "chair",
              "vector_query": "embedding:([], alpha: 0.8)",        
              "exclude_fields": "embedding"


When querying on both an embedding field and regular search fields, some parameters like query_by_weights won't have an impact on an embedding field mentioned in query_by. However, since length of query_by_weights must match the length of query_by, you can use a placeholder value like 0.

# Distance Threshold

You can also set a maximum vector distance threshold for results of semantic search and hybrid search. You should set distance_threshold in vector_query parameter for this.

# Hybrid Sorting

When doing a hybrid search (especially with a q parameter and an explicit vector_query parameter), you can sort by a combination of vector distance and also other numeric parameters using the special sort keyword _vector_distance in sort_by.

Here's an example:

  "sort_by": "popularity_score:desc,_vector_distance:asc"

Instead of combining the scores from both keyword and vector search, you can also use vector search distances as a sorting clause for reordering keyword search hits.

In the example below, we are using the vector distance as a secondary sorting condition to text match score.

  "q": "shoes",
  "query_by": "title,brand",
  "sort_by": "_text_match:desc,_vector_query(embedding:([0.43, 0.13, 0.21])):asc"

# Brute-force searching

By default, Typesense uses the built-in HNSW index to do approximate nearest neighbor vector searches. This scales well for large datasets. However, if you wish to bypass the HNSW index and do a flat / brute-force ranking of vectors, you can do that via the flat_search_cutoff parameter.

For example, if you wish to do brute-force vector search when a given query matches fewer than 20 documents, sending flat_search_cutoff=20 will bypass the HNSW index when the number of results found is less than 20.

Here's an example where we are filtering on the category field and asking the vector search to use direct flat searching if the number of results produced by the filtering operation is less than 20 results.

curl 'http://localhost:8108/multi_search' \
      -X POST \
      -d '{
            "searches": [
                "collection": "docs",
                "q": "*",
                "filter_by": "category:shoes",
                "vector_query": "embedding:([0.96826, 0.94, 0.39557, 0.306488], k:100, flat_search_cutoff: 20)" 
# Be sure to replace `embedding` with the name of the field that stores your embeddings.

# UI Examples

  • Here's (opens new window) a demo that shows you how to implement Hybrid Search (Semantic Search + Keyword Search + Filtering + Faceting) using Typesense's built-in embedding generation mechanism.

  • Here's (opens new window) a demo that shows you how to implement a "Find Similar" feature using Vector Search in an ecommerce store.

    Click on "Find Similar" below each product tile for notes on how to implement this.

  • Here's (opens new window) a demo that shows you how to implement Semantic Search, using an external embeddings API and Vector Search.

Last Updated: 5/24/2024, 3:35:39 PM