Tune Recall Rate

Public Preview

Zilliz Cloud introduces a search parameter level to allow users to balance search recall and performance. It also provides another search parameter, enable_recall_calculation, to give users the estimated recall rate of the current search. You can combine these two parameters to tune the recall rate of vector searches.

📘Notes

• This feature is in Public Preview and you have to use a cluster compatible with Milvus v2.5.4 and later or with Milvus v2.4.14 and later.

• This feature is compatible the API and SDKs of the following versions:

• Python: v2.5.4 and v2.4.14

• Java: v2.5.3 and v2.4.10

• Node.js: v2.5.2 and v2.4.10

• RESTful: v2.5.4

Overview

The recall rate in Zilliz Cloud usually refers to the proportion of relevant results successfully retrieved by a search. It measures the system's ability to recover all the relevant items from a collection.

To calculate a search's recall rate, you can divide the number of relevant items retrieved by the total number of applicable items that should be retrieved. For example, if a search retrieves 90 of 100 relevant items, the recall rate should be 0.9 or 90%.

A high recall rate usually indicates a more precise search result, which may be time-consuming. You may want to tune the recall rate to balance the precision and efficiency of vector searches.

Set up a search request

To set up a search request with tunable recall, you must include the level parameter inside the search parameters as follows:

query_vector = [0.3580376395471989, ..., 0.9029438446296592],

res = client.search(
    collection_name="quick_setup",
    data=[query_vector],
    limit=3, # The number of results to return
    search_params={
        "params": {
            "level": 1 # The precision control
        }
    }
)

The level parameter ranges from 1 to 10 and defaults to 1. The default value results in a recall rate of 90%, which is typically sufficient for most use cases.

For scenarios that require a high recall rate (99% or above), try setting the level parameter to an integer between 6 and 10. If search efficiency is not a concern, you can set this parameter to 10 to get the most precise results.

📘Notes

If the top-most level settings still do not suffice, contact Zilliz Cloud support.

Tune recall rate

Zilliz Cloud also introduces another search parameter named enable_recall_calculation to facilitate the tuning process. Setting this parameter to True indicates that Zilliz Cloud will estimate the recall rate of the current search and includes the estimated recall rate along with the search results.

query_vector = [0.3580376395471989, ..., 0.9029438446296592],

res = client.search(
    collection_name="quick_setup",
    data=[query_vector],
    limit=3, # The number of results to return
    search_params={
        "params": {
            "level": 6 # The precision control
            "enable_recall_calculation": True # Ask for recall rate calculation
        }
    }
)

With the above search request, you can get an estimated recall rate of the current search as follows:

# data: [...], recalls: [0.98]

During the estimation process, Zilliz Cloud:

1. Searches with the level parameter set to the user-defined value, and

2. Conducts another search with the level parameter set to 10.

3. Use the second search with the level parameter set to 10 as the ground truth to estimate the recall rate.

While setting enable_recall_calculation to True, you can adjust the value of the level parameter to obtain multiple recall rates. By considering these estimated figures and the duration of each search, you can roughly estimate the appropriate level setting.

Limits

Currently, this feature is available only for Zilliz Cloud clusters in basic vector searches, filtered searches, and range searches.