Amazon Kendra is AWS's machine-learning enterprise search service, long used as the retrieval layer for AI agents and RAG chatbots. In 2026 that changed: Kendra entered maintenance mode on 30 June 2026 and closed to new customers on 30 July 2026, and AWS now recommends Amazon Bedrock Managed Knowledge Base with Amazon Bedrock AgentCore for new agent knowledge retrieval. Existing Kendra users stay supported but should plan a migration.
Updated October 2026
What changed in 2026
- Kendra is in maintenance mode. According to the Amazon Kendra availability change notice, there is no new feature development from 30 June 2026, and the service is closed to new customers from 30 July 2026. Existing customers keep bug fixes and security updates.
- Bedrock Agents became "Classic". Amazon Bedrock Agents is now Bedrock Agents Classic, closed to new accounts from 30 July 2026; AWS recommends Amazon Bedrock AgentCore instead. Knowledge Bases and Guardrails are not affected.
- AgentCore gained knowledge layers. In June 2026 AWS announced a generally available Managed Knowledge Base and Web Search for AgentCore, exposed to agents through AgentCore Gateway (AWS Insider).
- Agents are scaling. McKinsey's 2026 State of AI survey found about two in ten respondents say their organisations have scaled AI agents enterprise-wide, so retrieval choices now affect production systems, not pilots.
Why AI agents need a retrieval layer
An LLM on its own knows only its training data. An enterprise agent must answer from policies, manuals, tickets and contracts that change every week and that not every user may see. Retrieval-augmented generation (RAG) solves this: the agent searches approved content, adds the best passages to its prompt and answers with citations. For the underlying language techniques, see NLP techniques.
Kendra was popular for this job because it combined natural-language search, many native connectors (SharePoint, Confluence, S3, databases) and document-level access control in one managed service.
The classic Kendra + agent architecture (2023–2025)
- User asks a question in a chat app.
- Identity is checked (for example with Amazon Cognito) so results respect permissions.
- Optional PII redaction on the query.
- The agent calls Kendra's Retrieve API for relevant passages.
- An orchestration framework (LangChain, Bedrock Agents) adds passages to the prompt.
- The LLM writes a grounded answer with sources.
This pattern still works for existing Kendra customers. The issue is that it will not get new features.
Kendra vs Bedrock Managed Knowledge Base
AWS's own migration notes list the main gaps you need to plan for:
| Area | Amazon Kendra | Bedrock Managed Knowledge Base |
|---|---|---|
| Status in 2026 | Maintenance mode; closed to new customers | Recommended path for new search and RAG apps |
| Native connectors | 32+ | 7 (per AWS docs at the time of writing) |
| Search mode | Semantic search, with keyword features | Always hybrid search |
| Answer generation | Retrieval only; you add the LLM | Built-in Retrieve and Generate API |
| Extras | Query suggestions, facets, synonyms, spell check, feedback learning | No direct equivalents; AWS documents workarounds |
| Metadata filters | Includes startsWith / stringContains | Those operators not supported; schemas may need changes |
The 2026 reference architecture for AWS agents
- Agent runtime: AgentCore managed harness (declare model, tools and instructions) or code-defined agents using Strands, LangChain, OpenAI Agents SDK or Claude Agent SDK.
- Knowledge: Bedrock Managed Knowledge Base, connected to the agent through AgentCore Gateway as an MCP tool.
- Guardrails: Bedrock Guardrails on the model plus policy enforcement at the gateway.
- Memory and identity: AgentCore short- and long-term memory, scoped per user.
- Observability: end-to-end tracing of every retrieval and tool call.
A practical migration plan
- Inventory: list Kendra indexes, data sources, connectors, filters, synonyms and features in use.
- Check connector coverage: for sources with no native connector, plan an ingestion job into S3 or a custom pipeline.
- Rebuild access control: confirm document permissions carry over before any user sees results.
- Run both side by side: AWS recommends comparing relevance, latency and completeness in parallel before switching traffic.
- Build a golden question set: real user questions with known correct answers, scored before and after. Our guide to ML model testing explains how.
- Move the agent layer: if you also use Bedrock Agents Classic, migrate to AgentCore at the same time.
Use cases that still matter
- Financial services: policy and compliance Q&A with strict document permissions.
- Healthcare: treatment guidelines and claims rules for staff, with human review.
- Public sector: citizen and staff questions answered from official guidance, escalating to people when unsure.
- Industrial operations: SOPs, maintenance manuals and incident reports searchable by technicians. More examples in NLP applications.
Cost and governance checks before you switch
Kendra is priced mainly per index, while knowledge base and agent costs are driven by storage, embeddings, retrieval calls and model tokens. Model your real query volume before migrating, and test with production-like traffic. On governance, confirm three things: document-level permissions are enforced at retrieval time, every answer logs which sources it used, and personal data is redacted before it reaches the model. These checks matter more than raw relevance scores when an agent's answer can trigger an action.
Where it's heading: 2026 → 2030
| Period | Enterprise knowledge retrieval |
|---|---|
| 2020 (history) | Enterprise search engines return ranked documents |
| 2023–2025 | Kendra and vector stores feed RAG chatbots |
| 2026 (now) | Managed knowledge bases exposed to agents as MCP tools; Kendra in maintenance |
| 2028 (expected) | Gartner expects 33% of enterprise software applications to include agentic AI (Gartner), each needing governed retrieval |
| 2030 (our view) | Retrieval spans clouds and on-prem stores with one policy layer; agents cite every source |
How NexaStack helps
Not every knowledge base can live in one cloud. NexaStack lets teams run agents, embedding models and retrieval services on their own infrastructure, on-premises, at the edge or across clouds, while still using AWS where it fits. That gives you a portable agent stack that does not depend on a single managed search service. ElixirData, the decision harness for enterprise AI agents, adds policy and audit for answers that drive decisions. For the wider tooling picture, see ML platform tools.
FAQ
What does Amazon Kendra do?
Kendra is a managed enterprise search service that uses machine learning to answer natural-language questions across connected document sources, with access control.
Is Amazon Kendra being discontinued?
It is in maintenance mode, not shut down. Since 30 July 2026 it is closed to new customers. Existing customers remain supported with bug fixes and security updates, but no new features.
What replaces Amazon Kendra?
AWS recommends Amazon Bedrock Managed Knowledge Base for new search and RAG applications, typically used with Amazon Bedrock AgentCore for agents.
What is the difference between Amazon Comprehend and Kendra?
Comprehend analyses text (entities, sentiment, classification). Kendra searches and retrieves content. Agents often used both: Comprehend to redact PII, Kendra to retrieve passages.
What are the latest AI agent retrieval trends in 2026?
Managed knowledge bases exposed through MCP, hybrid search by default, per-user memory, end-to-end tracing and running retrieval close to private data.
Build a portable knowledge layer for your agents
Planning a Kendra migration or running agents beyond one cloud? Book a NexaStack demo to see private retrieval and agent runtimes on your own infrastructure.
