Core functionality for using AI with the Nuxeo Platform.
This repository provides 3 packages:
nuxeo-ai-core
: Contains the core interfaces and AI componentnuxeo-ai-image-quality
: Enrichment services that uses Sightengine.nuxeo-ai-aws
: Enrichment services that use Amazon Web Services.
Ai-core Version | Nuxeo Version |
---|---|
2.0.1 | 10.3 |
2.1.0 | 10.10 |
2.1.1 | 10.10-HF02 |
2.1.2 | 10.10-HF05 |
2.1.3 | 10.10-HF22 |
2.2.x | 10.10-HF23+ |
3.0.x | 11.1-SNAPSHOT |
- Install the nuxeo-ai-core package.
./bin/nuxeoctl mp-install nuxeo-ai-core
You can set these in your nuxeo.conf
.
Parameter | Description | Default value | Since |
---|---|---|---|
nuxeo.ai.export.batch.size |
Sets batch size for the TF Record export | 20 |
Since 2.1 |
nuxeo.ai.export.bucket.size |
Sets bucket size for the TF Record export | 100 |
Since 2.1 |
nuxeo.ai.export.training.batch.size |
Sets how many records are getting into the training writer | 200 |
Since 2.1 |
nuxeo.ai.conversion.rendition |
Rendition title for multi-picture views to return for predictions/exports | Small |
Since 2.3 |
nuxeo.ai.conversion.strict |
Use Nuxeo renditions only during export | true |
Since 2.5 |
It is recommended that the Elasticsearch mappings are updated to allow a full text search on enrichment labels. The following code will add this mapping to a server running locally.
curl -X PUT \
http://localhost:9200/nuxeo/_mapping/doc/ \
-H 'Cache-Control: no-cache' \
-H 'Content-Type: application/json' \
-d '{
"dynamic_templates": [
{
"no_enriched_raw_template": {
"path_match": "enrichment:items.raw.*",
"mapping": {
"index": false
}
}
},
{
"no_enriched_norms_template": {
"path_match": "enrichment:items.normalized.*",
"mapping": {
"index": false
}
}
},
{
"no_enriched_history_template": {
"path_match": "enrichment:history.*",
"mapping": {
"index": false
}
}
}
],
"properties": {
"enrichment:items": {
"properties": {
"model": {
"type": "keyword",
"ignore_above": 256
},
"inputProperties": {
"type": "keyword",
"ignore_above": 256
},
"suggestions": {
"properties": {
"labels": {
"properties": {
"confidence": {
"type": "float"
},
"label": {
"type": "keyword",
"copy_to": [
"all_field"
],
"ignore_above": 256,
"fields": {
"fulltext": {
"analyzer": "fulltext",
"type": "text"
}
}
}
}
},
"property": {
"type": "keyword",
"ignore_above": 256
}
}
}
}
},
"enrichment:filled": {
"type": "keyword",
"ignore_above": 256
},
"enrichment:corrected": {
"type": "keyword",
"ignore_above": 256
}
}
}'
You can set these in your nuxeo.conf
.
Parameter | Description | Default value | Since |
---|---|---|---|
nuxeo.ai.export.min.docs |
Set minimum amount of documents required for continues expoort | 10 |
Since 2.1 |
nuxeo.ai.images.enabled |
Create a stream for creation/modification of images. | false |
Since 1.0 |
nuxeo.ai.video.enabled |
Create a stream for creation/modification of video files. | false |
Since 1.0 |
nuxeo.ai.audio.enabled |
Create a stream for creation/modification of audio files. | false |
Since 1.0 |
nuxeo.ai.text.enabled |
Create a stream for text extracted from blobs. | false |
Since 1.0 |
nuxeo.ai.stream.config.name |
The name of the stream log config | pipes |
Since 1.0 |
nuxeo.enrichment.source.stream |
The name of the stream that receives Enrichment data | enrichment-in |
Since 1.0 |
nuxeo.enrichment.save.tags |
Should enrichment labels be saved as a standard Nuxeo tags? | false |
Since 1.0 |
nuxeo.enrichment.save.facets |
Should enrichment data be saved as a document facet? | true |
Since 1.0 |
nuxeo.enrichment.raiseEvent |
Should an `enrichmentMetadataCreated` event be raised when new enrichment data is added to the stream? | true |
Since 1.0 |
nuxeo.ai.default.threshold |
Default Threshold value. Should be a float type between 0.0 and 1.0 | 0.75 |
Since 1.0 |
nuxeo.ai.autofill.default.threshold |
Default Threshold value for autofill. Should be a float type between 0.0 and 1.0 | 0.75 |
Since 1.0 |
nuxeo.ai.autocorrect.default.threshold |
Default Threshold value for autocorrect. Should be a float type between 0.0 and 1.0 | 0.75 |
Since 1.0 |
Nuxeo AI Core provides 3 Java modules:
- nuxeo-ai-core - Contains the core interfaces and AI component
- nuxeo-ai-pipes - Nuxeo Pipes, short for "Pipelines" provides the ability to operate with Nuxeo Stream. Nuxeo Stream provides a Log storage abstraction and a Stream processing pattern. Nuxeo Stream has implementations with Chronicle Queues or Apache Kafka.
- nuxeo-ai-model - Adds support for custom machine learning models
- Provides an
AIComponent
to register services. eg. An enrichment service. - Interfaces and helper classes for building services.
- Provides a
EnrichingStreamProcessor
to act on a stream using an JavaEnrichmentProvider
. - An Operation called
EnrichmentOp
to call anEnrichmentProvider
and return the result. - Provides a
RestClient
andRestEnrichmentProvider
for easily calling a custom json rest api. - Provides a
ThresholdComponents
to register type/facet based thresholds.
<extension target="org.nuxeo.ai.configuration.ThresholdComponent"
point="thresholdConfiguration">
<thresholdConfiguration type="Document"
global="0.8">
<thresholds>
<threshold xpath="dc:title"
value="0.6"
autofill="0.65"
autocorrect="0.70"/>
</thresholds>
</thresholdConfiguration>
</extension>
-
Continues Export - to enable such you must contribute a cron contribution to your instance. It has to fire
startContinuousExport
at a give time. For instance to export weekly you might use the following example:<extension target="org.nuxeo.ecm.core.scheduler.SchedulerService" point="schedule"> <schedule id="continuous_export_default"> <event>startContinuousExport</event> <!-- At 03:00 AM, every 7 days --> <cronExpression>0 0 3 */7 * ?</cronExpression> </schedule> </extension>
On such event the system will retrieve all AI Model defined under configured project and perform evaluation of the current data to check if it needs to be uploaded.
- Enables sending custom events to a nuxeo stream.
- Provides a
FunctionStreamProcessorTopology
to act on a stream using a JavaFunction<T, R>
. - Provides 4 customizable document streams:
images
- When a image is added to a document.videos
- When a video is added to a document.audio
- When an audio file is added to a document.text
- When binary text is extracted from a document.
These streams are disabled by default but can be enabled by the corresponding configuration parameters.
The configuration parameters are used to configure Nuxeo xml contributions, instead you can provide your own configuration that meets your requirements.
A Sample DAM configuration is available to download , and defines 2 pipelines:
- Listens for
pictureViewsGenerationDone
and sendspicture:views/3/content
to theimages
stream. - Configures an
EnrichingStreamProcessor
to read from theimages
stream, calls theaws.celebrityDetection
enrichment service and puts the response in theai/images-enrichment-in
stream. - The next stream processor reads from the
ai/images-enrichment-in
stream, and raises animageMetadataCreated
event for each new enrichment entry. - An example listener for the
imageMetadataCreated
event writes a log message.
- Listens for new
vid:storyboard
modifications for a document in a path containingmovies
and sends 4 of the video storyboard images to thevideo
stream. - Configures an
EnrichingStreamProcessor
to read from thevideo
stream, calls theaws.imageLabels
enrichment service and puts the response in theai/video-enrichment-in
stream. - The next stream processor reads from the
ai/video-enrichment-in
stream and creates document tags for the enrichment labels.
Please note that the EnrichingStreamProcessors
are using a stream processing policy of continueOnFailure=true
, this
means that stream processing will continue even if the enrichment failed.
Using an Nuxeo extension you can dynamically register a pipeline for any custom event.
For example to send MY_EVENT
to a stream called mystream
you would use the following configuration.
<extension point="pipes" target="org.nuxeo.ai.Pipeline">
<pipe id="pipe.mypipe" enabled="true" function="org.nuxeo.my.DocumentPipeFunction">
<supplier>
<event name="MY_EVENT">
<filter class="org.nuxeo.ai.pipes.filters.NotSystemOrProxyFilter"/>
</event>
</supplier>
<consumer>
<stream name="mystream"/>
</consumer>
</pipe>
</extension>
Transforming an input Event into an output stream is done using a function specified by the function
parameter.
Functions are explained below.
To enable/disable Nuxeo Insight Cloud enrichers based on your custom model set nuxeo.ai.insight.enrichment.enabled
to
the desired value. Default value is true
New enrichment services can be added by implementing EnrichmentProvider
. AbstractEnrichmentProvider
is a good
starting point. If you wish to call a custom rest api then extending RestEnrichmentProvider
would allow access to the
various RestClient
helper methods. To register your extension you would use configuration similar to this.
<extension point="enrichment" target="org.nuxeo.ai.services.AIComponent">
<enrichment name="custom1" kind="/classification/custom"
class="org.nuxeo.ai.custom.CustomModelEnrichmentProvider" maxSize="10000000">
<option name="minConfidence">0.75</option>
</enrichment>
</extension>
Actions on streams or events are based on the standard Java Function<T, R>
interface. To send an event to a stream you
would need to implement the Function<Event, Record>
interface. Record
is the type used for items in a nuxeo-stream.
For examples, look at PropertiesToStream
and its helper class DocEventToStream
.
These is also a FilterFunction
that first tests a Predicate
before applying the function. Predicates can be built
with the help of the Predicates
class. To create a predicate for only document events with documents which are not
system documents or proxies and aren't "Folderish" you would use this
predicate: docEvent(notSystem().and(d -> !d.hasFacet("Folderish"))
.
Stream processing is achieved using a computation stream pattern that enables you to compose producers/consumers into a complex topology.
To use a custom processor, create a class that implements FunctionStreamProcessorTopology
and specify it in
the class
parameter as shown below.
<extension target="org.nuxeo.runtime.stream.service" point="streamProcessor">
<streamProcessor name="basicProcessor" defaultConcurrency="1" defaultPartitions="4"
class="org.nuxeo.my.custom.StreamProcessor">
<option name="source">ai/mystream</option>
<option name="sink">ai/mystream-out</option>
</streamProcessor>
</extension>
You can register your custom enrichment services to act as a stream processor using the EnrichingStreamProcessor
. For
example, the following configuration would register a stream processor that acts on a source stream called images
, it
runs the custom1
enrichment service on each record and sends the result to the ai/enrichment-in
stream.
<extension target="org.nuxeo.runtime.stream.service" point="streamProcessor">
<streamProcessor name="myCustomProcessor1" defaultConcurrency="2" defaultPartitions="4"
class="org.nuxeo.ai.enrichment.EnrichingStreamProcessor">
<option name="source">ai/images</option>
<option name="sink">ai/enrichment-in</option>
<option name="enrichmentProviderName">custom1</option>
</streamProcessor>
</extension>
Nuxeo AI adds additional metrics to the standard Nuxeo Metrics reporting.
Metric name | Metric |
---|---|
nuxeo.ai.streams.[eventListener].events |
Count of events received. |
nuxeo.ai.streams.[eventListener].consumed |
Count of events that matched the filter condition and were processed. |
nuxeo.ai.enrichment.[enrichmentProvider].called |
Count of stream records received. |
nuxeo.ai.enrichment.[enrichmentProvider].errors |
Count of errors. |
nuxeo.ai.enrichment.[enrichmentProvider].produced |
How many records were produced after calling the service. |
nuxeo.ai.enrichment.[enrichmentProvider].retries |
Count of retries. |
nuxeo.ai.enrichment.[enrichmentProvider].cacheHit |
Count of times the result was returned from the cache rather than calling the enrichment service. |
nuxeo.ai.enrichment.[enrichmentProvider].unsupported |
Count of unprocessable records, perhaps due to mime-type or size. |
nuxeo.ai.enrichment.[enrichmentProvider].success |
Count of successful calls. |
nuxeo.ai.enrichment.[enrichmentProvider].circuitbreaker |
Incremented when the circuilt breaker is open, stopping the stream from any more processing. |
nuxeo.ai.enrichment.[enrichmentProvider].fatal |
Incremented when a fatal error occurs stopping the stream from any more processing. |
nuxeo.ai.streams.func.[functionName].called |
Count of stream records received. |
nuxeo.ai.streams.func.[functionName].errors |
Count of errors. |
nuxeo.ai.streams.func.[functionName].produced |
How many records were produced by the function. |
- When using the Chronicle implementation of nuxeo-stream you should make sure your
defaultPartitons
setting for stream processors matches the number of partitions you have, eg. 4.
Edit $NUXEO_HOME/lib/log4j2.xml
, in the <Appenders>
section, add a AI-FILE
appender:
<RollingFile name="AI-FILE" fileName="${sys:nuxeo.log.dir}/nuxeo-ai.log"
filePattern="${sys:nuxeo.log.dir}/nuxeo-ai-%d{yyyy-MM-dd}.log.gz" append="true">
<PatternLayout pattern="%d{ISO8601} %-5p [%t] [%c] %m%n"/>
<CronTriggeringPolicy schedule="0 0 0 * * ?" evaluateOnStartup="true"/> <!-- Rollover at midnight every day -->
<DefaultRolloverStrategy/>
</RollingFile>
Then in the <Loggers>
section, add a logger pointing to the AI-FILE
appender:
<Logger name="org.nuxeo.ai" level="debug">
<AppenderRef ref="AI-FILE"/>
</Logger>
Nuxeo Stream is either implemented with Chronicle Queues or Apache Kafka. To watch the progress of messages in the stream you can use:
$NUXEO_HOME/bin/stream.sh --help
For example, to see the last 8 messages in the "images" stream, for chronicle you would use the first command below (
passing in --chronicle nxserver/data/stream/pipes
) and for Kafka you would use the second command below (passing in
just -k
).
./bin/stream.sh tail -n 8 --chronicle nxserver/data/stream/pipes -l images --codec avro
./bin/stream.sh tail -n 8 -k -l images --data-size 2000 --codec avro
Similarly, to view the consumer lag on the "images" stream, for chronicle use the first command, and the second for kafka. The response format is Markdown.
./bin/stream.sh lag --chronicle nxserver/data/stream/pipes -l images --verbose
./bin/stream.sh lag -k -l images --verbose
When a document is enriched with suggestions, the suggestions are stored in the Enrichment
facet. It is possible to
automatically set a property with a suggested value in two ways. If the property currently has no value then you can
use Autofill
, if the property already has a value then use Autocorrect
.
- If the target property is
null
or its previously been autofilled then attempt to set the property. - Calculate the highest ranking suggestion from the suggestions stored in the "Enrichment" facet.
- If its confidence is greater than the threshold returned from the
ThresholdService
then set the property value.
- If its confidence is greater than the threshold returned from the
- If we have already autofilled this property but it doesn't meet the threshold then reset it to
null
. - Raise an
AUTO_FILLED
document event.
- If its previously autofilled then don't attempt an autocorrect because autofill and autocorrect are mutually exclusive.
- Calculate the highest ranking suggestion from the suggestions stored in the "Enrichment" facet.
- If its confidence is greater than the threshold returned from the
ThresholdService
then set the property value. - Save the previous value in the history blob (unless it was previously auto-corrected).
- If its confidence is greater than the threshold returned from the
- If we have already autocorrected this property but it doesn't meet the threshold then reset it to the previous value from the history.
- Raise an
AUTO_CORRECTED
document event.
Dataset exports use the Bulk Action Framework. To track progress of your bulk action you can use a command like this:
export COMMAND_ID=your-bulk-action-id
curl -s -X GET "localhost:8080/nuxeo/api/v1/bulk/$COMMAND_ID" -u Administrator:Administrator -H 'content-type: application/json'
The documentation on Debugging The Bulk Action Framework has more useful stream commands. Some further examples are:
./bin/stream.sh lag --chronicle /var/lib/nuxeo/data/stream/bulk -l ai/bulkDatasetExport
./bin/stream.sh lag --chronicle /var/lib/nuxeo/data/stream/bulk -l ai/writing
./bin/stream.sh tail -n 8 --chronicle /var/lib/nuxeo/data/stream/bulk -l done --codec avro --schema-store /var/lib/nuxeo/data/avro/ --data-size 3000
./bin/stream.sh tail -n 8 --chronicle /var/lib/nuxeo/data/stream/bulk -l command --codec avro --schema-store /var/lib/nuxeo/data/avro/ --data-size 3000
Nuxeo dramatically improves how content-based applications are built, managed and deployed, making customers more agile, innovative and successful. Nuxeo provides a next generation, enterprise ready platform for building traditional and cutting-edge content oriented applications. Combining a powerful application development environment with SaaS-based tools and a modular architecture, the Nuxeo Platform and Products provide clear business value to some of the most recognizable brands including Verizon, Electronic Arts, Netflix, Sharp, FICO, the U.S. Navy, and Boeing. Nuxeo is headquartered in New York and Paris. More information is available at www.nuxeo.com.