Natural language & keyword support
Amazon Kendra's ability to understand natural language questions is at the core of its search engine, so end users have the ability to search for general keywords like "health benefits" or more specific natural language questions like "how long is maternity leave?”. Kendra will return specific answers like "14 weeks", or for the more general searches Kendra will return the most relevant passage and related documents. Natural language enables you to get more specific answers from anywhere in your data.
Reading comprehension & FAQ matching
Amazon Kendra can extract specific answers from unstructured data. No pre-training is required, you simply point Kendra at your content and Kendra will provide specific answers to natural language queries like "how do I configure my VPN?" where the answer is automatically extracted from the most relevant document. You can also upload a list of FAQs to Kendra to provide direct answers to common questions your end users are asking. Kendra will find the closest question to the search query and return the corresponding answer.
To complement the extracted answers, Amazon Kendra uses a deep learning based semantic search model to return a ranked list of relevant documents. This provides the end user a more exhaustive list of content to explore should they need more information.
Use Kendra's connectors for popular sources like S3, SharePoint, Salesforce, Servicenow, RDS databases, One Drive and many more coming later this year. Using connectors is quick and easy, you just add data sources to your Kendra index and select the connector type. Connectors will maintain document access rights and can be scheduled to automatically sync your index with your data source, so you're always securely searching through the most up to date content. For other data source types, Kendra offers an API that allows you to build your own connector and upload documents from your ETL job or back-end application. Because each data source may contain different file types, Kendra supports unstructured and semi-structured data in HTML, MS Office Word and PowerPoint, PDF, and text formats.
You can boost certain fields in your index in order to assign more importance to specific responses. Amazon Kendra allows you to tune for specific data sources or document freshness. For example, when searching for “When is re:invent?” you can boost the relevance of document freshness so that the 2019 dates are the suggested answer. Or, you could boost a more reputable data source in an index of research reports. Kendra also supports boosting documents based on a vote or view count, which is common in forums and other support type knowledge bases. Combining these boosting features would, for example, boost documents that are not only viewed more often but that are also more recent, like trending news or updates.
With confidence scores, you can see the estimated accuracy of search results. Amazon Kendra assigns confidence scores based on four main categories: Very High, High, Medium, and Low for various document results (e.g. Factoid Answer, FAQ). Confidence scores enable you to distinguish and set thresholds for what results should be displayed.
Kendra uses deep learning models to understand natural language queries and document content and structures for a wide range of internal use cases like HR, operations, support, and R&D. Kendra is also optimized to understand complex language from domains like IT, financial services, insurance, pharmaceuticals, industrial manufacturing, oil and gas, legal, media and entertainment, travel and hospitality, health, HR, news, telecommunications, mining, food and beverage, and automotive. For example, a user searching for HR answers could enter "deadline for filing HSA form" and Kendra would also search for "deadline for filing health savings account form" for broader coverage to get the most accurate answer. You can further improve domain expertise by providing your own synonym lists (coming soon). You just upload a file with your specific terminology and Kendra will use these synonyms to enrich user searches.
Incremental learning coming soon
Amazon Kendra's machine learning models are regularly retrained and tuned for each customer by capturing end-user search patterns and feedback. For example, when users search for “How do I change my health benefits?", multiple HR and benefit documents will compete for a top spot. Although more recent documents can be ranked higher in the list, this criteria alone might not be sufficient to win the top spot on the list. However, as interest grows for the latest year’s enrollment instructions and users start opening it more than the older ones, the Kendra model will learn to increasingly prioritize the new document towards the top of the list.
Query auto-completion coming soon
Amazon Kendra includes the functionality to auto-complete an end user’s search query. Query auto-completion not only helps the user reduce typing by about 25%, it also helps them by guiding them towards more precise and commonly asked questions. More precise questions typically result in more relevant and useful answers. For example, if you start typing "Where is" in the search box, Kendra can suggest options like "Where is the IT desk?", or "Where is the cafeteria?" and other related commonly asked questions, to complete the query.
Analytics and ongoing improvement coming soon
In order to provide a continuously improving search experience, Amazon Kendra captures the activities during searches (clicks, thumbs up or down) and will surface metrics and insights to you so that you can take action to make improvements. Kendra provides basic operational metrics like top queries, top documents, and queries per day. Kendra will also provide common quality metrics such as Mean Reciprocal Rank (MRR) and explicit feedback, i.e. thumbs up or down counts. You can use this information to create more relevant FAQs, boost certain data sources with more authoritative content, or train your customer support teams on commonly asked questions.