OpenRAG

OpenRAG is a modular framework for building systems that answer questions using a document collection as their basis. It combines semantic retrieval with BM25 keyword search and multilingual reranking, and provides citations to the source document and page with links to the cited location. Supported document handling includes PDF layout awareness, OCR, image captioning, and audio transcription, across listed text, office, email, audio, video, and image formats. Users can connect their own models or a hosted provider. An admin console, chat interface, and OpenAI-compatible API are included, alongside integrations such as Open WebUI, LangChain, n8n, and Twake.ai. Partition isolation, owner/editor/viewer roles, and token or OpenID Connect authentication provide access controls. OpenRAG runs on infrastructure controlled by the deployer, keeping documents, embeddings, and queries within that perimeter. Its AGPL-3.0 self-hosted plan costs nothing. Docker, Docker Compose, and at least 16 GB RAM are listed as prerequisites; a CPU-only profile is optional. Several capabilities, including CSV support and encryption in transit and at rest, are listed as coming soon.

OpenRAG brings document retrieval, cited answers, and deployment on the operator’s infrastructure into one framework. Review the listed prerequisites and coming-soon items to see whether its current capabilities fit your needs.

OpenRAG may suit organizations building document-based assistants, legal search, or multimodal enterprise question-answering systems. It is relevant to teams that want to deploy on their own infrastructure and manage knowledge-base access by role.

OpenRAG is a self-hostable framework for building document-grounded search and AI assistants. It suits organizations that need control over where their knowledge and queries reside and can operate the required infrastructure. Its strongest draw is traceable, multimodal retrieval; teams wanting a turnkey service or capabilities still in development should look elsewhere.

Overview

OpenRAG combines semantic search, BM25 keyword matching and multilingual reranking, then returns answers with citations to the source document and page. The linked citation opens at that page, giving users a direct way to check an answer—particularly valuable in legal search and enterprise knowledge work, where an unsupported response is harder to trust.

The AGPL-3.0 framework runs on infrastructure controlled by the deployer, keeping documents, embeddings and queries within that perimeter. That control comes with an operational cost: Docker and Docker Compose are prerequisites, and the stated minimum is 16 GB of RAM. An optional CPU-only profile broadens hardware choices, but does not remove the need to deploy and maintain the system.

Key features

Retrieval and document processing

OpenRAG handles text, Markdown, PDF, DOC and DOCX, PPTX, email, audio, video and image files. Its multimodal parsing includes PDF layout awareness, OCR, image captioning and audio transcription, which makes it relevant to collections that extend beyond ordinary text documents. CSV, ODT and HTML support and format-specific chunkers are still coming soon, so teams whose current archives rely on those formats should not assume they are supported.

Interfaces, models and scaling

An admin console, chat interface and OpenAI-compatible API provide several ways to work with the system. Users can connect their own models—including Mistral, Qwen, Lucie, Claude and GPT—or use a hosted provider. Ray distributes ingestion, chunking and embedding across worker nodes for horizontal scaling, a useful option as collections grow, though it entails operating the deployment that supports those workers. Named integrations include Open WebUI, LangChain, n8n and Twake.ai.

Controls and security

Partitions isolate knowledge bases, with owner, editor and viewer roles; API tokens are stored as SHA-256 hashes. The security design also includes fail-closed scopes, verified outbound connections, redacted secrets, non-root containers, rate limiting and security headers. Encryption in transit and at rest is still coming soon, as are tool calling, agentic RAG and MCP. Those gaps matter for teams that require encrypted transport and storage or depend on tool-enabled workflows.

OpenRAG can run on customer infrastructure or hosting qualified SecNumCloud. Traceable answers and auditable code provide evidence relevant to GDPR and the AI Act, but compliance obligations remain with the deployer; the software does not transfer that responsibility.

Pricing

The self-hosted OpenRAG plan costs nothing under AGPL-3.0, with deployment on your own infrastructure and no feature held back. There are no paid feature tiers to trade against the free plan, but the 16 GB RAM minimum and deployment work mean that a zero software price is not the same as a turnkey rollout.

A LINAGORA-managed service is also offered with custom pricing. Billing can be usage-based or fixed-fee depending on deployment mode, with short, renewable commitment periods. This is the option to consider if you want a managed service rather than running the framework yourself; the self-hosted plan is the clearer fit for teams prepared to operate their own infrastructure.

Platforms

OpenRAG is available through an API, a web interface and self-hosted deployment. Its infrastructure-controlled model is a practical fit for teams that need documents and queries to stay within their own perimeter, but it is not a substitute for checking whether your team can meet the runtime requirements.

Who it's for

OpenRAG is aimed at teams building AI assistants, legal search and multimodal enterprise question answering. Public administrations and private companies use it. It is best suited to organizations that value source-linked answers, need access controls over knowledge bases and can manage a self-hosted system. For technical questions, feature suggestions and bug reports, the maker directs users to GitHub Issues.

Pros and cons

• Pro: Answers cite a document and page, with links opening at the cited page, making verification more direct.
• Pro: Hybrid retrieval, reranking and multimodal parsing cover mixed-format knowledge collections, including audio and images.
• Pro: Self-hosting keeps documents, embeddings and queries within the deployer's infrastructure; partitions and role controls help separate access.
• Con: Docker, Docker Compose and at least 16 GB of RAM make deployment a meaningful infrastructure commitment.
• Con: CSV, ODT and HTML support, encryption in transit and at rest, and tool-oriented capabilities remain in development.
• Con: Compliance responsibilities stay with the deployer, even when the hosting and code provide relevant evidence.

Alternatives

OpenRAG belongs among Retrieval-Augmented Generation Tools. Consider AnythingLLM if Android, desktop platforms or a freemium option matters alongside self-hosting; its listed Basic plan is $50 a month. Graphiti is another free, self-managed framework, with an Apache-2.0 license rather than OpenRAG's AGPL-3.0.

PrivateGPT is another free, open-source option for self-hosted use. Choose Quivr if its listed free EUR plans for knowledge management and support-data analysis better match the task. txtai is a free Apache 2.0 alternative installable with pip or Docker. Dify offers a freemium hosted or self-hosted route, with a Sandbox plan that includes 200 message credits, one workspace, one member, five apps and 50 knowledge documents. Haystack Open Source Framework is a free Python framework; Haystack Enterprise Platform is the alternative to consider for a platform with a free Studio tier and cloud deployment.

Verdict

Choose OpenRAG if your team needs cited, multimodal retrieval and wants documents, embeddings and queries kept on infrastructure it controls. Its combination of hybrid search, access controls and a free AGPL-3.0 self-hosted plan is compelling for organizations able to run it. Look elsewhere if you need a turnkey deployment, encryption in transit and at rest today, or file and agent capabilities that remain on the roadmap.