At the in-house exposition re:invent AWS presented its new machine learning based search service Amazon Kendra. The previous article was initially dedicated to natural voice control – and presented Alexa for Business as a suitable interface. Now we turn our attention to Kendra itself.
Part 2: „Kendra, fill me in!”
Amazon itself describes Kendra as a “highly accurate and easy to use enterprise search service that’s powered by machine learning”. Essentially, the idea is to provide a powerful search engine for distributed knowledge and internal documents that can be used with natural language.
Now the difficulty with the internal company search, in contrast to the web search with Google, for example, is that the data is often unstructured and located in different, sometimes isolated systems. In addition, documents often use different vocabulary and are available in different formats.
Kendra offers a solution for each of these problems. Data is made available to Kendra via so-called connectors. Amazon has already developed connectors for file systems, websites, DropBox, Salesforce, SharePoint, relational databases and Amazon S3. Other data sources can be connected via an API. Kendra supports unstructured and semi-structured data in HTML, Microsoft Word and PowerPoint, PDF as well as plain text formats.
Kendra understands what it’s all about…
Because every industry comes with its own language, Kendra is currently optimized for 16 different industries, including key sectors such as telecommunications, automotive and many more.
In addition, Kendra uses deep learning models to understand the structure, content of documents and – in contrast to a classical search based on the frequency of a search term – the context of the question. A search for ” What is the vacation policy?” will therefore produce results that address the topic without explicitly mentioning the keywords.
…and improves constantly
Using Kendra offers numerous advantages. Most fundamentally, by enabling employees and users to search the contents of manuals, reports and documentation, FAQs and more in a targeted manner and to answer internal company questions independently.
The service supports a wide range of data sources and simplifies the retrieval of information by resolving common abbreviations, for example, and also combining them with industry-specific knowledge – so that UPS, for example, stands for “Uninterruptible Power Supply” and is not recognized for the parcel service provider or “WTF” as the World Taekwondo Federation instead of this… uhm, you know.
In addition, the search service becomes more intelligent the longer it is used. Thus, Kendra enables the analysis of the questions asked in order to uncover possible knowledge gaps or further potential for improvement and to continuously improve the quality of the results, as its machine learning algorithms are constantly learning.
Currently Amazon Kendra is still in the preview phase. For instance, not all connectors have yet been implemented and the service is not yet available in all regions. The rollout is announced for 2020.
Ben Freiberg is a Senior Data Engineer and proven AWS expert at the *um location and AWS Competence Center in Frankfurt.
This post is also available in: German