How to replatform Endeca rules to Solr
Apr 07, 2020 • 8 min read
Apr 07, 2020 • 8 min read
Retail is on the brink of yet another technological breakthrough. AI devices like Siri, Alexa, Google Home, and virtual assistants of all kinds are rapidly becoming trusted and dependable aides to many people. Technology companies led by Amazon, Google, Apple, and Microsoft are rushing to grab the market of AI devices, applications, and platforms.
Each of these tech giants is hurriedly trying to build an AI ecosystem around its devices that is integrated with its backend cloud. This will set up the latest, and most likely largest ever, platform war for conversational commerce. The prize is control over consumers’ hearts, minds, and wallets in the brave new world of digital assistant-curated Internet experiences.
In a lot of ways, the arrival of voice interfaces with AI capabilities resembles the early days of the mobile revolution. This was a time when many tech companies were fiercely fighting for dominance in the emerging mobile market by pushing forward and promoting their platforms. All of this happened while retailers tried to wrap their heads around how to reach their customers through this new mobile channel and iron out the bumps and glitches of their immature mobile platforms. Those who persisted on this journey eventually became the successful omnichannel retailers of the mobile age.
Today’s retailers wanting to deliver their services to consumers via new AI channels face similar challenges of handling a new platform, except it’s now on an uneven playing field. The mobile age was driven by software and device companies that were impartial to the retail vertical so that the playing field was level for both small and large retail players. This is not the case with today’s AI platforms.
Take Alexa for example. Alexa is not an impartial AI device that would be equally happy to sell items from either Amazon or its competitors. Instead, it is designed to be strongly preferential to Amazon services and is deeply integrated with them. You want to check the delivery status of an item sold on Amazon? Sure thing! You want to check the delivery status of an item sold on Acme.com? I don’t think so...
In addition, targeting each device individually with the associated unique ecosystem of services is expensive to build and maintain. This is especially true when the number of AI devices on the market proliferates beyond two or three major players. This can be seen with the IM conversational channels such as Facebook Messenger, WhatsApp, and WeChat platforms.
So how can retail level the playing field and keep costs under control for AI? When the mobile revolution happened and mobile platforms started to diverge, web technology reacted with responsive/adaptive frameworks that allowed the building of device agnostic, omni-channel applications. We believe that the same should happen with AI and conversational commerce. What needs to emerge is an omni-channel CUI platform based on open source technologies that will treat Google Home, Amazon Echo, Messenger, and other upcoming voice and IM agents as end points that support voice-to-text and text-to-voice conversions.
The rest of the conversational system capabilities such as natural language understanding, dialog management, action fulfillment, etc. should work the same way across all channels so that the morning dialog with Amazon Echo can be picked up in the afternoon through Google Assistant and finished in the evening via Facebook Messenger. These capabilities can be backed by in-house implementations or be connected to the ecosystems of third-party conversational services.
At Grid Dynamics Labs, we are developing this vision with Grid Genie, an experimental open omni-channel conversational commerce platform. We are designing Grid Genie to support a wide range of conversational applications. The feature set will include the ability to guide the customer through product selection and associated research, facilitate purchases and returns, keep track of order shipments, and more.
Here is the conceptual architecture of the CUI platform:
There are quite a few components in the diagram, so let’s take a look at them one by one:
Let’s look at this architecture in action based on a simple example: order tracking.
One of the first use cases we addressed with the Grid Genie architecture is order tracking. Order tracking can be considered a good example of a relatively simple infobot experience. This is characterized by “shallow” conversation, when it is possible for the customer to specify what she is looking for in one question and receive all the information she needs in one answer. Here is an infographic of how this works.
However, even for such a simple conversation to work seamlessly across multiple devices, a lot of things need to happen under the hood. First and foremost, device-specific identity has to be mapped to customer identity within the store so that the conversational agent knows who it is talking to. Next, the customer’s voice needs to be accurately converted to text. Luckily, modern Speech2text models deployed by Google Assistant and Alexa are maturing quickly and are already very good.
Once the text, known as utterance, is received, we need to understand its broader intent. Is it a small talk attempt? Is it related to order tracking? Or product discovery? Maybe it is an answer to the previous question? Or is it just genuine gibberish? The dialog manager needs to perform intent classification and delegate the utterance to the appropriate agent.
Let’s assume a customer has asked an AI device about the shipping status of the Fisher-Price toy that she ordered the day before. After routing by the dialog manager, this utterance reaches the order tracking agent. It now has to perform natural language analysis and named entity extraction to understand what action is expected (search for an order), and what exactly the customer is looking for (a Fisher-Price toy ordered yesterday).
The agent is invoking a platform service to search for a Fisher-Price toy ordered yesterday which generates corresponding results.
Often the agent has to go back-and-forth with the search service on the customer’s behalf if they aren’t satisfied for whatever reason. Once satisfactory results are received from the order search service, the agent has to decide which template to use for the natural language response. This decision can be made via rules depending on the order results, or by running a predictive model to rank all available templates with this particular result.
Natural language responses have to be constructed to fill the slots in a selected response template with order data. Finally, the responses have to be converted back to text, taking into account the target device’s capabilities. If we are dealing with a voice agent, we also have to include speech synthesis markup to add more color and intonation to the response, or provide phonemes for unusual words such as brand names. If we are working in an IM channel, adding pictures or richer content may also be required.
As you can see, even a relatively simple infobot contains a lot of interacting components that have to work seamlessly together to achieve a natural conversational experience. This includes support for dialog context when the customer may be referring to a previous result (“when will it be delivered again?”), or omit specification of the intent (“what about my dress?”). Here, the conversational agent needs to understand what the customer is referring to. Check out our Grid Genie order tracking in action in this demo video.
Infobots, while already useful, are only just scratching the surface of the complexity of conversational channel. Fully-fledged conversational commerce must be able to deal with far more challenging “deep” conversations. This involves remembering large amounts of context dependent information that the customer has provided. This includes dealing with ambiguity of intent, asking questions to refine the customer’s needs, summarizing and analyzing large result sets, providing factoids about selected products, assisting in product comparisons, and coming up with relevant recommendations. In other words, it has to support the customer at every stage of the sales funnel - starting from discovery and ending with post-purchase support.
As you can see, this involves a huge amount of complexity, but we are working hard to teach Grid Genie new tricks every day. Stay tuned for more updates from Grid Labs!