Open Source AWS Kendra Alternatives

Elasticsearch is a distributed search and analytics engine, a scalable data store and a vector database, tuned for speed and relevance on production-scale workloads. It is the foundation of Elastic's open Stack platform and searches in near real time over very large datasets.

Data is reached through a REST API or one of the official language clients, which keeps it usable from any stack. Elastic publishes its machine learning work and its Lucene contributions through Search Labs, so the engine's direction is visible rather than opaque.

The same engine covers several jobs that are often bought separately.

• Full-text search: relevance ranking over documents at scale.
• Vector search: embeddings stored and queried alongside the rest of your data.
• Retrieval augmented generation: the retrieval half of generative AI applications.
• Observability: logs, metrics and application performance monitoring.
• Security analytics: security logs kept searchable for investigation.

You can create a managed deployment on Elastic Cloud or download and run it yourself. For a laptop, a start-local script brings up Elasticsearch and Kibana in Docker, generates a password and an API key, and writes them to an environment file. That setup leaves transport encryption off and binds to localhost, so it belongs in development only, and its one-month trial of all features reverts afterwards to the free and open Basic tier.

Rather than build chat, retrieval and agent plumbing in house, a team can stand up Onyx and get an application layer for large language models that it hosts itself. Company knowledge arrives through more than 50 indexing connectors or through MCP.

Model choice stays open. The platform works with self hosted runtimes such as Ollama, LiteLLM and vLLM as well as proprietary providers including Anthropic, OpenAI and Gemini.

Several capabilities sit on top of that foundation.

• Agentic RAG: a hybrid index paired with retrieval agents for search and answer quality.
• Deep research: multi step research flows that produce in depth reports.
• Custom agents: agents given their own instructions, knowledge and actions.
• Web search: an in house crawler plus support for Serper, Google PSE, Brave, SearXNG and others.
• Code execution: a sandbox for analysing data, rendering graphs or modifying files.

Deployment runs through Docker, Kubernetes, Helm or Terraform, with a single install command for a quick start. A Lite mode trims the stack to a chat interface under 1GB of memory, while the standard deployment adds the vector and keyword index, background job workers, model inference servers and caching for larger use.

The Community Edition is MIT licensed and covers chat, RAG, agents and actions, with an Enterprise Edition adding single sign on, role based access control, analytics, query history and whitelabeling.

PipesHub connects AI applications to the knowledge already scattered across a company's business systems, so an answer no longer depends on which tool somebody happened to search first. It links Slack, Google Drive, GitHub, Microsoft 365, Notion and more than fifty other systems into one context layer.

Retrieval runs over a knowledge graph that captures relationships between records, with block level citations pointing back at the original document. Any LLM provider can sit behind it, deployed inside your own VPC so data never leaves your infrastructure.

Several jobs sit on top of that shared context layer.

• Permission-aware search: source level access controls are enforced, so people only see records they are already authorized to read.
• Enterprise connectors: more than fifty connectors arrive with real time and scheduled indexing.
• No-code agents: agents are built visually and can execute actions across connected tools.
• Artifacts and code execution: reports, charts and dashboards are generated inside a sandbox.
• Multimodal input: images, diagrams and scanned files are understood, alongside voice interaction.

The backend is Python with FastAPI, storing documents in MongoDB, vectors in Qdrant, OpenSearch or Redis, and the graph in Neo4j or ArangoDB. Deployment is Docker Compose, driven by an interactive installer that checks prerequisites, generates secrets and writes an environment file for you. A managed cloud edition is announced but not yet available, so self-hosting is the only way to run it today.

Swirl answers a question by searching your applications live, ranking what comes back and returning an answer with sources you can click through. The premise is that most AI search products ask you to copy everything into a vector database first and then govern that copy forever, and that this step can be skipped.

Queries run against each source with the user's own permissions, so access control is enforced where it already lives rather than rebuilt in a second system. Results are re-ranked, deduplicated and mixed, then an answer with citations is generated by whichever model you supply a key for. Everything lands in SQLite or Postgres for later analysis.

This repository is Swirl Community, the Apache 2.0 edition.

• Connectors: more than a hundred sources are supported, and the connector objects are built to be extended.
• Galaxy UI: the same interface that ships with the commercial edition.
• Re-ranking: cosine vector similarity using the spaCy large model and NLTK, plus duplicate detection and result mixers.
• Processors: a pipelined architecture transforms queries, responses and results as they pass through.
• Federation: synchronous or asynchronous, over a clean REST API.

Getting started is a compose file and one Docker command, roughly two minutes, arriving ready to search Arxiv, European PMC and Google News. That Docker setup keeps nothing when it shuts down, so follow the persistent install guide for real use. A paid enterprise edition adds a three-pass reranker, canonical answers and an MCP server for agents.

Corpus takes PDFs and web pages, answers questions about them in plain language, and links each answer back to the passage it came from. Documents group into workspaces and document sets, so a research pile stays separated by project instead of merging into one searchable heap.

The backend is FastAPI with PostgreSQL as the primary database, Elasticsearch for full-text search, RabbitMQ as the message queue, Temporal for workflow orchestration, Redis for caching and S3 for document storage. The frontend is React with Vite. A Docker Compose file starts the whole stack, including a LocalStack S3 and the Temporal web interface.

The product sits on a handful of moving parts.

• Citations: every answer carries links back into the source text.
• Cross-correlation: the same question can be compared across several documents at once.
• Workflows: Temporal runs automated workflows whose output can be reviewed and exported.
• Model choice: OpenAI, Anthropic, Google and xAI models are supported for answering.
• Embeddings: OpenAI and Voyage AI both work for the vector side.

Running everything in containers needs only Docker and Docker Compose, while running the services directly needs Python 3.11 or newer with Poetry and Alembic, plus Node.js 20 for the frontend. Video walkthroughs in the repository cover getting started, cross-correlation and workflow creation, and the license is AGPL-3.0.