What Is Amazon Kendra? Enterprise Search on AWS
Published on February 25, 2026
Your employees are spending 3.6 hours every single day hunting for documents buried inside SharePoint, Confluence, S3 buckets, and Salesforce — and finding nothing useful.
For a company with 500 knowledge workers at an average salary of $80,000, that is $14.4 million per year walking out the door in pure search friction.
That is not a productivity issue. That is an infrastructure decision with a measurable cost attached to it.
Amazon Kendra is Amazon Web Services' fully managed, ML-powered intelligent enterprise search service. It replaces keyword-based tools with Natural Language Processing that understands intent — not just terms — so employees get direct answers instead of a list of 200 loosely related documents.
We have deployed Kendra-based search architectures for enterprises across the US, UAE, and Singapore. Here is exactly what it does, where it wins, and where it will disappoint you if you are not careful.
The Real Problem With Your Search Stack Right Now
Here is what is actually happening inside your organization right now. A compliance officer types "parental leave California employees" into SharePoint Search. SharePoint returns 237 documents containing those words — ranked by term frequency. She opens 14 of them before finding the actual answer buried in Policy HR-2024-07.
That Same Search in Amazon Kendra: 9 Seconds
Kendra reads the question, understands the intent, scans your indexed documents, and returns: "Employees in California receive 12 weeks of paid parental leave under Policy HR-2024-07" — with the source document linked. Done.
First-result accuracy: from 23% to 78%
Verified with a 3,000-employee healthcare client across 45,000+ documents
We constantly see enterprises running Elasticsearch or SharePoint Search as their "intelligent" enterprise search layer when they are, frankly, just glorified CTRL+F across documents. Those tools were built for keyword matching. Kendra was built for questions.
How Amazon Kendra Actually Works (Not the Sales Version)
Kendra's architecture has four core pieces. Know them, or your deployment fails at step 2.
The 4 Core Architecture Pieces
1. The Index
The central brain. Kendra ingests documents, processes them, generates embeddings, and stores search-ready data. Developer Edition ($810/month, 10,000 docs, 4,000 queries/day) or Enterprise Edition ($1,008/month, 100,000 docs, 8,000 queries/day). Start Developer. Always.
2. Data Source Connectors
14+ native connectors: S3, SharePoint Online, Confluence, Salesforce, ServiceNow, OneDrive, Google Drive, Jira, and more. Each connector inherits Access Control Lists (ACLs) from the source system. No custom ETL pipelines needed.
3. Query Classification
Every query gets classified: factoid ("How many PTO days?"), descriptive ("Explain onboarding"), keyword ("expense report template"), navigational ("HR benefits portal"). Different types trigger different retrieval strategies. This is why Kendra beats Elasticsearch on natural language.
4. Result Layers
Three result types per query: a direct extracted answer, a highlighted document excerpt, and a ranked document list as fallback. Users get the best possible answer regardless of phrasing.
(The one thing AWS will not say in the sales deck: ACL sync happens during scheduled data source syncs — not in real-time. For sensitive roles with recently revoked access, set incremental syncs to every 1–2 hours minimum or you are sitting on a security gap.)
Amazon Kendra Pricing: What It Actually Costs in 2026
Stop reading posts that quote 2022 pricing. Here is the real cost structure today:
| Tier | Monthly Base | Documents | Queries/Day | Best For |
|---|---|---|---|---|
| Developer Edition | $810/month | 10,000 | 4,000 | POC, teams under 50 users |
| Enterprise Edition | $1,008/month | 100,000 | 8,000 | Production deployments |
| Additional Documents | $0.000001/doc/hr | — | — | Scale beyond included limits |
| Additional Queries | $0.00034/query | — | — | Overage billing |
Real-World Cost Ranges
Mid-size (500 users, 50,000 docs, 5,000 queries/day): $2,500–$4,000/month. Large enterprise (5,000+ users, 500,000 docs): $8,000–$15,000/month.
Healthcare client deployment cost: $4,200/month. IT helpdesk tickets dropped 42% in Q1.
$9,200/month recovered in ticket resolution costs — deployment paid for itself in under 6 weeks
Here is the controversial opinion: Most companies do not need Enterprise Edition to start. Developer Edition handles 10,000 documents and 4,000 queries/day — that covers 90% of proof-of-concept deployments. Run on Developer for 60 days, measure business impact, then go to leadership with actual ROI data instead of vendor projections. AWS themselves report an 82% lower 5-year TCO versus traditional enterprise search tools.
Kendra vs. Elasticsearch vs. OpenSearch: The Honest Take
Everyone asks us this comparison, so here is the blunt answer.
Use Amazon Kendra When...
Employees ask questions in plain English and need precise answers from internal documents: HR portals, IT helpdesks, legal knowledge bases, customer support portals.
Use OpenSearch or Elasticsearch When...
You need sub-100ms latency on high-volume application search, log analytics, or product catalogs with complex filtering. Kendra's 1–3 second query latency and per-query pricing make it expensive and sluggish at 50,000+ queries/day.
The Architecture We Recommend
Kendra for employee Q&A + OpenSearch for application search. Two tools, two jobs, optimized cost. Kendra does not replace OpenSearch — it completes it.
Use Azure Cognitive Search only if your entire stack lives inside the Microsoft ecosystem and you have zero appetite for AWS.
Kendra + Amazon Bedrock: The RAG Architecture That Delivers
This is where enterprise search gets genuinely interesting in 2026.
Pair Amazon Kendra as your retrieval layer with Amazon Bedrock (Claude 3.5 Sonnet or Amazon Titan) as your generation layer, and you get a knowledge assistant that answers in natural language — grounded in your actual documents, without hallucinating facts. Kendra pulls the right documents. Bedrock synthesizes the answer from them.
Financial Services Client — 3,200+ Daily Queries
Before Kendra + Bedrock: agents averaged 4.7 minutes per ticket lookup. After implementation: 61% of queries handled automatically in under 12 seconds. Human-handled tickets dropped average handle time from 4.7 to 2.1 minutes.
Deployment took 11 working days. ROI hit in week 3.
Security and Compliance: What Gets Ignored at Deployment
Amazon Kendra covers the compliance checklist: SOC 1/2/3, ISO 27001, HIPAA eligible, PCI DSS Level 1, FedRAMP Authorized (Moderate), GDPR compliant, and CCPA compliant. For healthcare organizations, Kendra can be included in your AWS Business Associate Agreement.
The #1 Cause of Failed Kendra Deployments
IAM role misconfiguration. Kendra requires separate IAM roles for the index, each data source connector, and FAQ uploads. Teams that use one catch-all role hit either security holes or permission errors at data source sync time.
Use least-privilege roles. Create a separate IAM role per data source. Never attach AdministratorAccess to any Kendra role in production. (Your auditors will not ask twice about this one.)
Stop Losing $14.4 Million a Year to Search Friction
Your knowledge workers are not unproductive. They are spending 3.6 hours every day searching because your search infrastructure is a keyword box that was never designed for natural language. At Braincuber Technologies, our AWS team deploys Kendra-based intelligent search for enterprises in healthcare, financial services, and manufacturing — handling index architecture, connector setup, ACL mapping, relevance tuning, and full RAG integration.
Frequently Asked Questions
What file types can Amazon Kendra index?
Kendra indexes PDFs, Word (.docx), PowerPoint (.pptx), HTML, CSV, JSON, plain text, and structured FAQ lists. Individual documents are capped at 5 MB each. Encrypted or password-protected files need pre-processing via AWS Lambda through Kendra's document enrichment pipeline before they can be indexed.
How long does a Kendra deployment actually take?
A basic Developer Edition setup — one index, 2–3 connectors (SharePoint + S3 + Confluence), and 200 FAQ entries — takes 1–2 weeks. Adding a custom React or Next.js search UI with relevance tuning adds another 1–2 weeks. A full Enterprise Edition deployment with RAG integration using Amazon Bedrock runs 3–6 weeks end-to-end.
Does Kendra work with on-premises data sources?
Yes, through custom connectors built via AWS Lambda. Native connectors are cloud-based (SharePoint Online, Salesforce, S3), but on-premises SharePoint Server, SQL Server, or internal wikis connect through Kendra's custom data source API with Lambda as the bridge. Incremental sync schedules keep your index current automatically.
What is the difference between Amazon Kendra and Amazon Q Business?
Amazon Q Business is a newer, higher-level generative AI assistant that uses Kendra-style retrieval combined with LLM generation out of the box — faster to deploy for broad employee productivity use cases. Amazon Kendra gives developers full control over index architecture, relevance tuning, and retrieval logic — better for custom applications where search behavior must be precisely controlled.
How many languages does Amazon Kendra support?
Kendra supports 15+ languages: English, Spanish, French, German, Japanese, Portuguese, Chinese (Simplified and Traditional), Italian, Korean, Arabic, and others. Language is defined at the data source level. Cross-language query support within a single index requires custom pre-processing — Kendra does not auto-translate queries between languages natively.