Intelligent Search - Cross-Searching Internal Knowledge with Amazon Kendra
Learn how to build enterprise search with Amazon Kendra. Covers natural language queries, data source connectors, RAG (Retrieval-Augmented Generation) integration, and when to choose Kendra vs. OpenSearch.
Enterprise Search Challenges and Where Kendra Fits
Enterprise knowledge is scattered across many systems, such as S3, SharePoint, Confluence, internal wikis, FAQ sites, and databases. Traditional keyword search requires users to know the right search terms and cannot handle natural language questions such as "How do I apply for paid leave?" Amazon Kendra is an ML-based intelligent search service that comes standard with data source connectors for major business systems, understands natural language questions, and extracts answers directly from documents and returns them. Language support has a two-tier structure: in addition to English (the default language), 34 languages are covered by keyword matching, and of these, 8 languages (English, Japanese, Chinese (Simplified and Traditional), French, German, Korean, Spanish, and Portuguese (Brazil)) support full semantic search, including answer extraction through reading comprehension and confidence scores. For example, for the question "By when do new employees need to take their health checkup?", Kendra identifies the relevant section in an internal policy PDF and presents the answer. While OpenSearch (Elasticsearch-based) is strong at keyword matching with inverted indexes and at aggregation and analysis, Kendra specializes in semantic search powered by natural language understanding (NLU) and machine learning. OpenSearch, which engineers use for log analysis and metrics search, and Kendra, which non-engineer business users use to search internal knowledge, complement each other. However, Kendra stopped accepting new customers on July 30, 2026 (use by existing customers continues, as of August 2026), so for new builds from now on, Amazon Bedrock Knowledge Bases is the official alternative.
Data Source Connectors and Index Building
Kendra provides data source connectors for major SaaS and on-premises products as standard, and connections to representative systems are completed without code. Supported sources include S3 (PDF, Word, HTML, text), RDS/Aurora (text columns in databases), SharePoint Online, Confluence, Salesforce, ServiceNow, Google Drive, OneDrive, Slack, and GitHub. With the custom connector API, you can also integrate your own data sources. The default quota for the number of data source connectors registered to a single index is 5 for Basic Developer Edition and 50 for the Enterprise editions, and the number of indexes itself has a default quota of 10 per account (you can request increases for the number of indexes and for the Enterprise connector count, but the Developer limit of 5 cannot be raised). Data source synchronization runs automatically on a schedule (hourly, daily, weekly), and new, updated, and deleted documents are reflected in the index. With the feature that automatically retrieves ACLs (access control lists) from data sources, access permissions in SharePoint and Confluence are reflected as is in Kendra search results. Documents that User A does not have permission to view do not appear in the search results. This transparent access control is an important feature in enterprise environments with strict security requirements.
Search Features and RAG Integration
Kendra returns search results in three types. Suggested Answer extracts a direct answer from a document with the relevant passage highlighted. FAQ Answer returns answers from a pre-registered FAQ database. Document returns a list of highly relevant documents, similar to traditional search results. For search tuning, you can configure field boosting (weighting specific metadata fields), faceted search (filtering by category, date, author), and synonym dictionaries (thesaurus). Kendra can serve as a retriever in Bedrock's RAG (Retrieval-Augmented Generation) pattern. Kendra searches for relevant documents in response to a user's question, then passes those documents as context to a Bedrock LLM, generating accurate answers grounded in internal knowledge. Kendra's high-precision search results directly improve RAG answer quality.
Pricing and Edition Selection
(Figures are as of August 2026 for the Tokyo Region, with monthly amounts based on 730 hours) Kendra offers three types of indexes. Basic Developer Edition costs 1.125 USD per hour (about 821 USD per month) and supports 10,000 documents or 3 GB of extracted text and 0.05 queries per second (about 4,000 queries per day). It is suited to development and test environments and small-scale internal search, but because it does not support purchasing additional storage units or query units, you need to move to a higher-tier index when you run out of capacity. Basic Enterprise Edition costs 1.40 USD per hour (about 1,022 USD per month), supports 100,000 documents or 30 GB of extracted text and 0.1 queries per second (about 8,000 queries per day), and also provides high availability (Multi-AZ). If capacity is insufficient, you can add storage units and query units at 0.70 USD per hour each, up to 100 units. The edition officially recommended for use as a RAG retriever is GenAI Enterprise Edition, whose base price is much lower at 0.32 USD per hour; it starts with 20,000 documents or 200 MB of extracted text and scales with storage units (0.25 USD per hour) and query units (0.07 USD per hour). However, it is available only in US East (Northern Virginia), US West (Oregon), Europe (Ireland), and Asia Pacific (Sydney), and cannot be selected in the Tokyo Region. What cost design most often overlooks is how to end the charges. Charges begin as soon as you create an index, and you pay even when you use no storage or query capacity at all. There is no operation equivalent to pausing, and the only way to end the charges is to delete the index. Shutting indexes down at night or on holidays to save money is not possible, so build your plans for test environments on the assumption that you delete the whole index when you finish using it. In addition, the Basic editions (Enterprise and Developer) charge, separately from the base price, 0.35 USD per hour of connector run time and 0.000001 USD per document scanned for data source synchronization. The connector fee for GenAI Enterprise Edition is a flat 30 USD per month per index, which includes up to 500 hours of synchronization per month. Compared with OpenSearch Serverless, Kendra excels in natural language search accuracy, but its cost is higher. A reasonable way to divide the roles is OpenSearch when the search targets are structured log data and Kendra when they are unstructured internal documents.
Tuning Search Accuracy and Continuous Improvement
Enterprise search does not end with deployment; it is something whose accuracy you keep improving. What to grasp before anything else here is the language constraint. The three features of synonym dictionary (thesaurus) registration, incremental learning that reflects user feedback, and query suggestions that complete search keywords apply only to indexes in English, the default language. If Japanese internal documents are your main target, design on the assumption that these features are unavailable. The tuning methods available in Japanese are weighting by boosting metadata fields, faceted filtering by category, date, or author, prioritized answers to routine questions by registering FAQs, and organizing titles and headings on the document side. Analyze the search logs, and when you find patterns of questions that do not return the desired results, work out which method will fill each gap. The decision goes like this: field boosting if a particular department's documents do not rank near the top, FAQs for frequent routine questions, and content-side work if no relevant document exists at all. If an index consists mainly of English documents, you can run a continuous improvement loop that also includes synonym registration and feedback. Continuing to refine the settings leads to a search experience that truly puts internal knowledge to use.
Summary - Guidelines for Using Kendra
Amazon Kendra is an intelligent search service that searches across internal knowledge using ML-based natural language understanding. Its main strengths are data source connectors for major business systems, transparent access control through ACLs, FAQ functionality, and RAG integration with Bedrock. It delivers results in situations where information retrieval is a business bottleneck, such as reducing internal helpdesk inquiries, making knowledge management more efficient, and improving the quality of customer support answers. If an existing customer sets up a new index, a straightforward approach is to validate on a small scale with Basic Developer Edition and, at production scale, move to Basic Enterprise Edition, which lets you add capacity units. If the available Regions match and you use it as a RAG retriever, GenAI Enterprise Edition, with its low base price, is the official recommendation. For new builds from now on, as mentioned at the beginning, Amazon Bedrock Knowledge Bases is the alternative.
References (Official AWS Resources)
The primary sources for this page are the official AWS website and documentation. Check the official pages below for the latest specifications and pricing.
- Amazon Kendra official page
- Amazon Bedrock official page
- Amazon OpenSearch Service official page
- AWS Documentation (official)
If this page and the official documentation disagree, treat the official documentation as authoritative.