When Google open-sourced TensorFlow, an end-to-end open source platform for machine learning, many onlookers immediately imagined the possibilities for improving their own search implementations with Google’s software framework. After all, Google was the largest consumer search company and owed much of its recent successes to artificial intelligence. A few of the bravest and most innovative companies rushed to lean on TensorFlow for a myriad of use cases. Lots of people use TensorFlow for image classification so it is a natural choice for adding machine learning to image search.

And while Incorporating TensorFlow into search applications was definitely an architecture that companies have explored, it was eventually abandoned because incorporating it in search proved to be one helluva challenge. Adding fields at index time or query time, is one frustrating example.

Fast forward to less than two years after TensorFlow was completely open-sourced. Lucidworks has released a version of Fusion that makes incorporating TensorFlow as simple as creating a model, training the model, bundling the resulting protobuf and a configuration file that will guide the TensorFlow graph’s serialization, and dropping it into an index pipeline.

Classify Documents and Queries With TensorFlow

You can now use Lucidworks Fusion to quickly deploy TensorFlow models for use in your search index or query pipelines. Currently, the simplest implementation is to use TensorFlow to add a field generated by machine learning to a document corpus.

You can use TensorFlow for classifying documents. While Fusion users could have classified documents through ML stages, custom JavaScript stages or other methods of classification, TensorFlow within Fusion allows users to take advantage of the flexibility, popularity, and sheer ubiquity of TensorFlow. If you have data scientists in your organization using TensorFlow to taxonomize your data, you may be able to repurpose those models to improve your search performance.

But, wait, there’s more. You can use TensorFlow for classifying queries too. With our models or your own, we can help you to understand whether a querying customer on your website is searching to comparison shop, to buy a specific product, or to look at what’s new this season. When you understand the emotions behind a query, you can make a stronger appeal in the more relevant results you show the user.

Test Drive TensorFlow

If you would like to test drive search with TensorFlow, you have a few options to get going. You can upgrade your Fusion instance to 4.2.0, or you can checkout the Lab Financial Oracles, which is available in the Create Instances drop-down here. Comment on this post if you have questions about deploying TensorFlow models with Fusion. The sentiment values you will see in the app as facets were generated in the index pipeline when the article text from each document was run through TensorFlow. They are the output of running Yahoo Finance articles through a sentiment analysis model.

TensorFlow models are not the only models you can deploy, either. Since Fusion 4.0.0, we have supported deploying a number of other models. The implications for voice search, image search, and even the standard text search will make this feature one of the most popular features within Fusion.

Whether you were seeking to improve image search, revenue on commerce, or search performance, TensorFlow’s comprehensive, flexible ecosystem of tools, libraries and community resources is now a viable answer.