Loosely Coupled Record Automation

I want to automate the creation of records in Content Manager so that workflows can be initiated and other content can be captured. As records and workflows are processed by users within Content Manager I want to export that activity to a graph database. That graph database will highlight relationships between records that might not otherwise be easily determined. It can also aide in tracing lineage and relationships.

I few years ago I would have created a lightweight C# console app that implements the logic and directly integrates with Content Manager via the .Net or COM SDK. No longer though! Now I want to implement this using as many managed services and server-less components as possible.

This diagram depicts the final solution design…

CM Graph (2).png

What does each component do?

  • Cloud Scheduler — an online cron/task utility that is cheap and easy to use on GCP

  • Cloud Functions — light-weight, containerized bundles of code that can be versioned, deployed, and managed independently of the other components

  • Cloud PubSub — this is a message broker service that allows you to quickly integrate software components together. One system may publish to a topic. Other systems (0+) will subscribe to those topics

  • Service API — REST API end-point that enables integration over HTTP

  • Content Manager Event Server — custom .Net Event Processer Plugin that publishes new record meta-data and workflow state to a PubSub topic

  • Graph Database — enables searching via cypher query syntax (think social network graph) across complex relationships

Why use this approach?

  • Centralized — Putting the scheduler outside of the CM server makes it easier to monitor centrally

  • Separation of Concerns — Separating the “Check Website” logic from the “Saving to CM” logic enables us re-use the logic for other purposes

  • Asynchronous Processing — Putting PubSub between the functions let’s them react in real-time and independently of each other

  • Scaling — cloud functions and pubsub can scale horizontally to billions of calls

  • Error handling — when errors happen in a function we can redirect to an error topic for review (which could kick-off a workflow)

  • Language Freedom — I can use python, node, or Go for the cloud functions; or I can use .Net (via Cloud Run instead of as a Cloud Functions)

Overall this is a pretty simple undertaking. It will grow much more complex as time progresses, but for now I can get building!


Fetching the records

This is super easy with python! My source is a REST API that will contain a bunch of data about firms. For each retrieved firm I’ll publish a message to a topic. Multiple things could then subscribe to that topic and react to the message.

First we’ll create the topic…

2019-06-21_18-33-09.jpg

Next I write the logic in a python module…

import urllib.request as urllib2
import sys
import json
import requests
import gzip
import os
from google.cloud import pubsub
 
project_id = os.getenv('GOOGLE_CLOUD_PROJECT') if os.getenv('GOOGLE_CLOUD_PROJECT') else 'CM-DEV'
topic_name = os.getenv('GOOGLE_CLOUD_TOPIC') if os.getenv('GOOGLE_CLOUD_TOPIC') else 'new_firm'
 
def callback(message_future):
    # When timeout is unspecified, the exception method waits indefinitely.
    if message_future.exception(timeout=30):
        print('Publishing message on {} threw an Exception {}.'.format(
            topic_name, message_future.exception()))
    else:
        print(message_future.result())
 
def downloadFirms(args):
    request_headers = requests.utils.default_headers()
    request_headers.update({
        'Accept''text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,image/apng,*/*;q=0.8', 
        'Accept-Language''en-US,en;q=0.9', 
        'Cache-Control''max-age=0', 
        'User-Agent''Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:52.0) Gecko/20100101 Firefox/52.0' 
    })
    url = "https://api..."
    request = urllib2.Request(url, headers=request_headers)
    html = gzip.decompress(urllib2.urlopen(request).read()).decode('utf-8')
    data = json.loads(html)
    try:
        hits = data['hits']['hits']
        publisher = pubsub.PublisherClient()
        topic_path = publisher.topic_path(project_id, topic_name)
        for firm in hits:
            firm_data = firm['_source']
            firm_name = firm_data['firm_name']
            topic_message = json.dumps(firm_data).encode('utf-8')
            print(firm_name)
            msg = publisher.publish(topic_path, topic_message)
            msg.add_done_callback(callback)
    except Exception as exc:
        print('Error: ', exc)
    finally:
        print('Done!')
 
if __name__ == "__main__":
    downloadFirms()

Now that module can be placed into a cloud function…

2019-06-21_18-41-02.jpg

Don’t forget to pass in the run-time parameters so that the function can post to the correct topic in the correct project. You may change these during your testing process.

2019-06-21_18-42-55.jpg

With that saved we can now review the HTTP end-point address, which we’ll use when scheduling the routine download. Open the cloud function and click onto the trigger tab, then copy the URL to your clipboard.

2019-06-21_18-46-06.jpg

In the cloud scheduler we just need to determine the frequency and the URL of the cloud function. I’ll post an empty json object as it’s required by the scheduler (even though I won’t consume it directly within the cloud function).

2019-06-21_18-49-34.jpg

Next I need a cloud function that subscribes to the topic and does something with the data. For quick demonstration purposes I’ll just write the data out to the console (which will materialize in stackdriver as a log event).

2019-06-21_19-14-16.jpg

With that created I can now test the download function, which should result in new messages in the topic, and then new output in Stackdriver. I can also create log metrics based on the content of the log. For instance, I can create a metric for number of new firms, number of errors, or average runtime execution duration (cloud functions cap out in terms of their lifetime, so this is important to consider).

2019-06-21_19-17-43.jpg

Now I could just put the “create folder in CM” logic within my existing cloud function, but then I’m tightly-coupling the download of the firms to the registration of folders. That would limit the extent to which I can re-use code and cobble together new feature functionality. Tightly-coupled solutions are harder to maintain, support, and validate.

In the next post we’ll update the cloud function that pushes the firm into the Content Manager dataset!

Enriching Record Metadata via the Google Vision API

Many times the title of the uploaded file doesn't convey any real information.  We often ask users to supply additional terms, but we can also use machine learning models to automatically tag records.  This enhances the user's experience and provides more opportunities for search.  

faulkner.jpg
Automatically generated keywords, provided by the Vision API

Automatically generated keywords, provided by the Vision API

In the rest of the post I'll show how to build this plugin and integrate it with the Google Vision Api...


First things first, I created a solution within Visual Studio that contains one class library.  The library contains one class named Addin, which is derived from the TrimEventProcessorAddIn base class.  This is the minimum needed to be considered an "Event Processor Addin".

using HP.HPTRIM.SDK;
 
namespace CMRamble.EventProcessor.VisionApi
{
    public class Addin : TrimEventProcessorAddIn
    {
        public override void ProcessEvent(Database db, TrimEvent evt)
        {
        }
    }
}

Next I'll add a class library project with a skeleton method named AttachVisionLabelsAsTerms.  This method will be invoked by the Event Processor and will result in keywords being attached for a given record.  To do so it will call upon the Google Vision Api.  The event processor itself doesn't know anything about the Google Vision Api.

using HP.HPTRIM.SDK;
 
namespace CMRamble.VisionApi
{
    public static class RecordController
    {
        public static void AttachVisionLabelsAsTerms(Record rec)
        {
 
        }
    }
}

Before I can work with the Google Vision Api, I have to import the namespace via the NuGet package manager.

The online documentation provides this sample code that invokes the Api:

var image = Image.FromFile(filePath);
var client = ImageAnnotatorClient.Create();
var response = client.DetectLabels(image);
foreach (var annotation in response)
{
    if (annotation.Description != null)
        Console.WriteLine(annotation.Description);
}

I'll drop this into a new static method in my VisionApi class library.  To re-use the sample code I'll need to pass the file path into the method call and then return a list of labels.  I'll mark the method private so that it can't be directly called from the Event Processor Addin.

private static List<string> InvokeDetectLabels(string filePath)
{
    List<string> labels = new List<string>();
    var image = Image.FromFile(filePath);
    var client = ImageAnnotatorClient.Create();
    var response = client.DetectLabels(image);
    foreach (var annotation in response)
    {
        if (annotation.Description != null)
            labels.Add(annotation.Description);
    }
    return labels;
}

Now I can go back to my record controller and build-out the logic.  I'll need to extract the record to disk, invoke the new InvokeDetectLabels method, and work with the results.  Ultimately I should include error handling and logging, but for now this is sufficient.

public static void AttachVisionLabelsAsTerms(Record rec)
{
    // formulate local path names
    string fileName = $"{rec.Uri}.{rec.Extension}";
    string fileDirectory = $"{System.IO.Path.GetTempPath()}\\visionApi";
    string filePath = $"{fileDirectory}\\{fileName}";
    // create storage location on disk
    if (!System.IO.Directory.Exists(fileDirectory)) System.IO.Directory.CreateDirectory(fileDirectory);
    // extract the file
    if (!System.IO.File.Exists(filePath) ) rec.GetDocument(filePath, false"GoogleVisionApi", filePath);
    // get the labels
    List<string> labels = InvokeDetectLabels(filePath);
    // process the labels
    foreachvar label in labels )
    {
        AttachTerm(rec, label);
    }
    // clean-up my mess
    if (System.IO.File.Exists(filePath)) try { System.IO.File.Delete(filePath); } catch ( Exception ex ) { }
}

I'll also need to create a new method named "AttachTerm".  This method will take the label provided by google and attach a keyword (thesaurus term) for each.  If the term does not yet exist then it will create it.

private static void AttachTerm(Record rec, string label)
{
    // if record does not already contain keyword
    if ( !rec.Keywords.Contains(label) )
    {
        // fetch the keyword
        Keyword keyword = null;
        try { keyword = new HP.HPTRIM.SDK.Keyword(rec.Database, label); } catch ( Exception ex ) { }
        if (keyword == null)
        {
            // when it doesn't exist, create it
            keyword = new Keyword(rec.Database);
            keyword.Name = label;
            keyword.Save();
        }
        // attach it
        rec.AttachKeyword(keyword);
        rec.Save();
    }
}

Almost there!  Last step is to go back to the event processor add in and update it to use the record controller.  I'll also need to ensure I'm only calling the Vision API for supported image types and in certain circumstances.  After making those changes I'm left with the code shown below.

using System;
using HP.HPTRIM.SDK;
using CMRamble.VisionApi;
 
namespace CMRamble.EventProcessor.VisionApi
{
    public class Addin : TrimEventProcessorAddIn
    {
        public const string supportedExtensions = "png,jpg,jpeg,bmp";
        public override void ProcessEvent(Database db, TrimEvent evt)
        {
            switch (evt.EventType)
            {
                case Events.DocAttached:
                case Events.DocReplaced:
                    if ( evt.RelatedObjectType == BaseObjectTypes.Record )
                    {
                        InvokeVisionApi(new Record(db, evt.RelatedObjectUri));
                    }
                    break;
                default:
                    break;
            }
        }
 
        private void InvokeVisionApi(Record record)
        {
            if ( supportedExtensions.Contains(record.Extension.ToLower()) )
            {
                RecordController.AttachVisionLabelsAsTerms(record);
            }
        }
    }
}

Next I copied the compiled solution onto the workgroup server and registered the add-in via the Enterprise Studio. 

2018-05-19_7-57-09.png

 

Before I can test it though, I'll need to create a service account within google.  Once created I'll download the API key as a json file and place it onto the server.

The API requires that the path to the json file be referenced within an environment variable.  The file can be placed anywhere on the server that is accessible by the CM service account.  This is done within the system properties contained in the control panel.

2018-05-19_7-51-45.png

Woot woot!  I'm ready to test.  I should now be able to drop an image into the system and see some results!  I'll use the same image as provided within the documentation, so that I can ensure similar results.  

2018-05-19_8-16-19.png

Sweet!  Now I don't need to make users pick terms.... let the cloud do it for me!

Automating the generation of Tesseract OCR text renditions

Although IDOL will index the contents of PDF documents, it does not perform its' own OCR of the content (at least the OEM connector for CM does not).  In the JFK archives this means I can only search on the stamped annotation on each image.  Even if IDOL re-OCR'd documents, I can't easily extract the words it finds.  I need to do that when researching records, performing a retention analysis, culling keywords for a record hold, or writing scope notes for categorization purposes.  In the previous post I created a record addin that generated a plain text file that held OCR content from the tesseract engine.    

Moving forward I want to automate these OCR tasks.  For instance, anytime a new document is attached we should have a new OCR rendition generated.  I think it makes sense to take the solution from the previous post and add to it.  The event processor plugin I create should call the same logic as the client add-in.  If this approach works out, I can then add a ServiceAPI plugin to expose the same functionality into that framework.

So I took the code from the last post and added another C# class library.  I added one class that derived from the event processor addin class.  It required one method be implemented: ProcessEvent.  Within that method I check if the record is being reindex, the document has been replaced, the document has been attached, or a rendition has changed.  If so I called the methods from the TextExtractor library used in the previous post. 

using HP.HPTRIM.SDK;
using System;
using System.IO;
using System.Reflection;
 
namespace CMRamble.Ocr.EventProcessorAddin
{
    public class Addin : TrimEventProcessorAddIn
    {
        #region Event Processing
        public override void ProcessEvent(Database db, TrimEvent evt)
        {
            Record record = null;
            RecordRendition rendition;
            if (evt.ObjectType == BaseObjectTypes.Record)
            {
                switch (evt.EventType)
                {
                    case Events.ReindexWords:
                    case Events.DocReplaced:
                    case Events.DocAttached:
                    case Events.DocRenditionRemoved:
                        record = db.FindTrimObjectByUri(BaseObjectTypes.Record, evt.ObjectUri) as Record;
                        RecordController.UpdateOcrRendition(record, AssemblyDirectory);
                        break;
                    case Events.DocRenditionAdded:
                        record = db.FindTrimObjectByUri(BaseObjectTypes.Record, evt.ObjectUri) as Record;
                        var eventRendition = record.ChildRenditions.FindChildByUri(evt.RelatedObjectUri) as RecordRendition;
                        if ( eventRendition != null && eventRendition.TypeOfRendition == RenditionType.Original )
                        {   // if added an original
                            rendition = eventRendition;
                            RecordController.UpdateOcrRendition(record, rendition, Path.Combine(AssemblyDirectory, "tessdata\\"));
                        }
                        break;
                    default:
                        break;
                }
            }
        }
        #endregion
        public static string AssemblyDirectory
        {
            get
            {
                string codeBase = Assembly.GetExecutingAssembly().CodeBase;
                UriBuilder uri = new UriBuilder(codeBase);
                string path = Uri.UnescapeDataString(uri.Path);
                return Path.GetDirectoryName(path);
            }
        }
    }
}
 

Note that I created the AssemblyDirectory property so that the tesseract OCR path can be located correctly.  Since this is spawned from TRIMEvent.exe the executing directory is the installation path of Content Manager.  The tesseract language files are in a different location though.  To work around this I pass the AssemblyDirectory property into the TextExtractor.

I updated the UpdateOcrRendition method in the RecordController class so that it accepted the assemblypath.  If the assembly path is not passed then I default the value to the original value which is relative.  The record add-in can then be updated to match this approach.

2017-11-14_20-53-36.png

Within the TextExtractor class I added a parameter to the required method.  I could then pass it directly into the tesseract engine during instantiation.  

2017-11-14_20-56-41.png

If you expand upon this concept you can see how it's possible to use different languages or trainer data.  For now I need to go back and add one additional method.  In the event processor I reacted to when a new rendition was added, but I didn't implement the logic.  So I need to create a record controller method that works for renditions.

public static bool OcrRendition(Record record, RecordRendition sourceRendition, string tessData = @"./tessdata")
{
    bool success = false;
    string extractedFilePath = string.Empty;
    string ocrFilePath = string.Empty;
    try
    {
        // get a temp working location on disk
        var rootDirectory = Path.Combine(Path.GetTempPath(), "cmramble_ocr");
        if (!Directory.Exists(rootDirectory)) Directory.CreateDirectory(rootDirectory);
        // formulate file name to extract, delete if exists for some reason
        extractedFilePath = Path.Combine(rootDirectory, $"{sourceRendition.Uri}.{sourceRendition.Extension}");
        ocrFilePath = Path.Combine(rootDirectory, $"{sourceRendition.Uri}.txt");
        FileHelper.Delete(extractedFilePath);
        FileHelper.Delete(ocrFilePath);
        // fetch document
        var extract = sourceRendition.GetExtractDocument();
        extract.FileName = Path.GetFileName(extractedFilePath);
        extract.DoExtract(Path.GetDirectoryName(extractedFilePath), truefalse"");
        if (!String.IsNullOrWhiteSpace(extract.FileName) && File.Exists(extractedFilePath)) {
            ocrFilePath = TextExtractor.ExtractFromFile(extractedFilePath, tessData);
            // use record extension method that removes existing OCR rendition (if exists)
            record.AddOcrRendition(ocrFilePath);
            record.Save();
            success = true;
        }
    }
    catch (Exception ex)
    {
    }
    finally
    {
        FileHelper.Delete(extractedFilePath);
        FileHelper.Delete(ocrFilePath);
    }
    return success;
}

Duplicating code is never a great idea, I know.  This is just for fun though so I'm not going to stress about it.  Now I hit compile and then register my event processor addin, like shown below.

2017-11-14_21-09-31.png

I then enabled the configuration status and saved/deployed...

2017-11-14_21-10-24.png

Over in the client I removed the OCR rendition by using the custom button on my home ribbon...

2017-11-14_21-13-59.png

When I then monitor the event processor I can see somethings been queued!

2017-11-14_21-11-55.png

A few minutes later I've got a new OCR rendition attached.

2017-11-14_21-17-24.png

Progress!  Next thing I need to do is train tesseract.  Many of these records are typed and not handwritten.  That means I should be able to create a set of trainer data that improves the confidence of the OCR text.  Additionally, I'd like to be able to compare the results from the original PDF and the tesseract results.