Page tree
Skip to end of metadata
Go to start of metadata

The Periscope Result Ranker module uses neural networks to store Find Bar search results and rank them by relevance. It extends the Periscope modules.

The module learns user preferences to offer better result ranking for subsequent searches. By default, search result rankings are stored per user. You can change the configuration as necessary. For example, you can enable an individual ranking for each user or for selected users only. Additionally, you can configure the memory size of networks to mitigate possible memory consumption issues in large setups.

Installation

Maven is the easiest way to install the module. Add the following dependency to your webapp. All other necessary dependencies will be added automatically.

<dependency>
  <groupId>info.magnolia.periscope</groupId>
  <artifactId>magnolia-periscope-result-ranker</artifactId>
  <version>1.1.1</version>
</dependency>

Pre-built JARs are also available for download. See Installing a module for help.

Configuration

The module comes with the following default configuration:

periscope-result-ranker/src/main/resources/periscope-result-ranker/config.yaml
outputUnits: 10000
rankingNetworkStorageStrategy:
  class: info.magnolia.periscope.rank.ml.jcr.JcrUsernameNetworkStorageStrategy

Properties

outputUnits 

required,  default is  10000

The memory size of neural networks.

The result-ranking system requires memory (heap space) and disk space per unit for each user (local ranking) or instance (global ranking). You can adjust the size of the memory used per unit to mitigate possible memory consumption issues (see Result Ranker memory size).

rankingNetworkStorageStrategy 

required

The result-ranking memory strategy.

The default strategy stores result rankings per user. Other strategies are possible (see Result Ranker strategy).

To adjust the strategy, set the class property accordingly.

class 

required, default is info.magnolia.periscope.rank.ml.jcr.JcrUsernameNetworkStorageStrategy

Other possible values must be a subtype of info.magnolia.periscope.rank.ml.RankingNetworkStorageStrategy.

Understanding configuration to optimize memory footprint

The Periscope Result Ranker module creates a certain number of memory units. The total number of memory units depends on the Result Ranker strategy. The size of a single memory unit is based on the Result Ranker memory size.

Result Ranker strategy

You can set the Result Ranker strategy via the class property of the rankingNetworkStorageStrategy property.

User-based ranking

This default configuration enables result ranking for each user operating on the author instance. With this strategy, the module creates one neural network for each user. The memory footprint grows with every new user working on an author instance.

Class: info.magnolia.periscope.rank.ml.jcr.JcrUsernameNetworkStorageStrategy

Role-based ranking

With this strategy, only users with the role superuser  or ranker have local (per-user) ranking memory. Any other users work with the global (per-instance) ranking memory.

Class: info.magnolia.periscope.rank.ml.jcr.JcrUserRoleNetworkStorageStrategy

To keep the memory footprint for the Result Ranker on a minimum level, use the JcrUserRoleNetworkStorageStrategy class and do not assign the role ranker. Make sure that you have only a few users with the superuser role.

Custom ranking

You can develop your own custom result-ranking strategy. To do this, create a custom class that implements RankingNetworkStorageStrategy and set the class property in the configuration accordingly.

Result Ranker memory size

You can globally configure the memory size of neural networks. To do so, adjust the outputUnits value. The default memory size is 10000; this value represents the maximum number of Find Bar search results that can be stored and ranked in networks.

Large networks can store more results but use up more memory, while small networks consume less memory but might lose stored results to free up additional memory. Results are removed from networks based on a least-recently-used policy. This ensures that frequent results remain in memory irrespective of when they were added to the networks. You may want to configure the memory size of networks depending on the Result Ranker strategy.

When you change the outputUnits value (e.g. from 10000 to 1000), you need to clean up the JCR rankings workspace. Otherwise, an error will appear when you select a search result. This is because the networks loaded into memory were created using a configuration that no longer exists, rendering any stored results obsolete. When you change the outputUnits value, make sure that you delete any stored networks and log into the Magnolia instance again to regenerate networks using the new configuration (see Clearing Result Ranker memory).

Changing configuration

The Periscope Result Ranker module configuration resides in periscope-result-ranker/src/main/resources/periscope-result-ranker/config.yaml. The module is deployed as a JAR file, but you can change the configuration by one of the following means:

The configuration is read by the Resources module. Magnolia scans the following for a resource (in this particular order):

  • JCR resources workspace (hotfix)
  • Light module
  • Classpath (JAR file)

The Java bean representation of the selected resource is then stored in the registry. The object in the registry may get decorated if a decorator exists for that configuration.

Storing configuration in the JCR resources workspace is useful for hotfixes and patches. This way, an administrator can create an urgent fix and publish it in the Resource Files app. In the long run, we recommend using file-based decoration.

The configuration data is read on startup and after it has been changed. The actual data is stored in the module's configuration registry. You can look it up using the Definitions app in modules periscope-result-ranker.

If you change the outputUnits value, you must delete all existing network data (see Clearing Result Ranker memory).

Changing configuration with resources hotfix

  1. Open the Resource Files app.
  2. Browse to and select periscope-result-ranker > config.yaml.
  3. In the action bar, click Edit file. The Resource Files app creates a copy of the currently used configuration and stores it in the JCR resources workspace.
  4. Edit the file as necessary.
  5. Click Save changes.

Changing configuration using decoration

With decoration, you can adapt the currently used configuration (whether it is from a light module, hotfix or JAR file). To learn more about decoration, see Definition decoration concept.

A decorator file can reside in any Magnolia Maven module or any light module (see Definition decorator file location). In the example below, we will create a decorator file in a light module named test-module.

Within the light module, create the file decorations/periscope-result-ranker/config.yaml.

<magnolia.resources.dir>/test-module/decorations/periscope-result-ranker/config.yaml
outputUnits: 1000
rankingNetworkStorageStrategy:
  class: info.magnolia.periscope.rank.ml.jcr.JcrUserRoleNetworkStorageStrategy

Clearing Result Ranker memory

The Periscope Result Ranker module stores all user-based, role-based, and custom rankings in the JCR rankings workspace. The module creates one node for each memory unit.

Nodes for local (per-user) rankings are named after user names. Nodes for global (per-instance) rankings are named default-neural-network-rankings.

To clear the Result Ranker memory:

  1. Open the JCR app.
  2. Switch to the rankings workspace.
  3. Select the nodes you want to delete.
  4. In the action bar, click Delete item.

Adding support for Linux armhf and ppc64le

The Periscope Result Ranker module supports the 64-bit versions of Linux, Mac OS and Windows by default. If you use Magnolia in one of the environments below, you must add the corresponding dependencies manually.

Linux armhf

<dependency>
 <groupId>org.bytedeco.javacpp-presets</groupId>
 <artifactId>openblas</artifactId>
 <classifier>linux-armhf</classifier>
</dependency>
Linux ppc64le
<dependency>
 <groupId>org.bytedeco.javacpp-presets</groupId>
 <artifactId>openblas</artifactId>
 <classifier>linux-ppc64le</classifier>
</dependency>
<dependency>
 <groupId>org.nd4j</groupId>
 <artifactId>nd4j-native</artifactId> 
 <classifier>linux-ppc64le</classifier> 
</dependency>