RAG
The best RAG platforms in 2026: an honest comparison

Shekhar Kadyan
Co-founder

The best RAG platform depends on what you are building and whose data it has to read. Choose LlamaIndex or LangChain to build your own pipeline, RAGFlow or Dify for an open-source builder, Amazon Bedrock Knowledge Bases or Azure AI Search if you are committed to one cloud, Vectara or Contextual AI for RAG as a service, and Glean, Onyx or PipesHub for permission-aware search and answers across company apps.
These choices fall into three groups. A framework suits you if you want to assemble your own pipeline in code. A RAG engine suits you if you want a visual builder over files you upload. A managed or enterprise platform suits you if you need connectors to business apps and answers that respect each person's permissions.
Disclosure: This article is published by PipesHub, and PipesHub is one of the products compared. Competitor facts are dated and linked to vendors' own pages. The tables are ordered by category and then alphabetically, and we say where a rival is the better choice. If you spot an error, tell us and we will fix it.
Key takeaways
RAG (retrieval-augmented generation) means finding relevant passages in your documents and handing them to a large language model (LLM). The model's answer then rests on your data and not on its memory.
The term "RAG platform" covers three different kinds of product. There are frameworks you assemble, RAG engines you configure, and managed or enterprise platforms that also handle connectors, permissions and hosting.
For company data, the deciding question is permission-aware retrieval. This means the system only retrieves what the person asking is allowed to see in the original app. Many tools leave this work to you.
No public benchmark covers every product here, so test on your own data and include a test for permission leaks.
Licences differ. Dify adds conditions to Apache 2.0, and several open-source projects keep some features in a paid edition.
What is a RAG platform?
A RAG platform handles the retrieval half of retrieval-augmented generation, and often the generation half too. It works in five steps:
It ingests content, which means it reads your files and app data.
It chunks the content, which means it splits it into passages.
It indexes the passages. This usually means storing embeddings, which are numeric fingerprints of meaning, alongside a keyword index.
It retrieves the best passages for each question.
It asks an LLM to answer with citations, which are links back to the exact sources it used.
Products come in three kinds. Frameworks such as LlamaIndex, LangChain and Haystack are code libraries. You pick the parts and run them yourself.
RAG engines such as RAGFlow and Dify are apps you can host yourself. They give you a web interface for uploading documents and building chat apps.
Managed and enterprise platforms run the whole pipeline for you. This group includes Vectara, Contextual AI, Glean, Onyx, PipesHub, and the cloud services from AWS, Microsoft and Google. The enterprise ones add connectors, which sync content automatically from apps like Slack, Google Drive, Jira or SharePoint. They also enforce who can see what.

The products compared in this article, from most to least to build yourself. Cloud services are shown on their own because you still build the user experience on top of them.
Comparison table (as of November 2026)
In the two tables below, "Yes" means the vendor documents the feature. "Partial" means the feature exists with limits that are explained below. "You build it" means you get the pieces but not the finished feature. There is deliberately no overall score.
Four terms appear in the tables. MCP is the Model Context Protocol, an open standard that lets assistants such as Claude, ChatGPT or Cursor call outside tools. An agent is an LLM program that can plan steps and use tools on its own. An ACL (access control list) is the list of who may open a document in its source app. A VPC (virtual private cloud) is a private section of a cloud provider that only your company uses.
What each option is
Option | Kind | Licence | Self-host | Best for |
|---|---|---|---|---|
Haystack | Framework | Apache 2.0 | Yes | Production pipelines in Python with a European vendor behind them |
LangChain | Framework | MIT | Yes | Custom agents and the largest integration ecosystem |
LlamaIndex | Framework | MIT (LlamaCloud is paid) | Yes (the framework) | Document-heavy RAG with strong parsing |
Dify | RAG engine / app builder | Dify Open Source License (Apache 2.0 plus conditions) | Yes | Building LLM apps and workflows visually |
RAGFlow | RAG engine | Apache 2.0 | Yes | Hard documents such as scans, tables and complex PDFs |
Amazon Bedrock Knowledge Bases | Cloud service | Proprietary (AWS) | No (it runs in your AWS account) | Teams already building on AWS |
Azure AI Search | Cloud service | Proprietary (Microsoft) | No (it runs in your Azure tenant) | Teams already building on Azure and Microsoft 365 |
Contextual AI | Managed RAG platform | Proprietary | In your own VPC (Enterprise) | Managed RAG agents, or RAG components bought separately |
Glean | Enterprise search platform | Proprietary | Partial (Customer Hosted, managed by Glean inside your GCP or AWS account) | Large companies wanting a finished assistant with support |
Onyx | Enterprise search platform (open source) | MIT Community + paid Enterprise | Yes (air-gappable) | Open-source enterprise search with a large community |
PipesHub | Enterprise RAG platform (open source) | Apache 2.0 Community + paid Enterprise | Yes | Open-source, permission-aware context for people and agents |
Vectara | RAG as a service | Proprietary (open-source MCP server) | On VPC and on-prem plans | Hosted RAG API with hallucination controls |
What each option includes
Option | Business-app connectors | Permission-aware retrieval | Citations | Agents / MCP |
|---|---|---|---|---|
Haystack | Partial (file and storage integrations) | You build it | You build it | Agents, plus MCP through MCPTool in the mcp-haystack package |
LangChain | Partial (many document loaders, which read once) | You build it | You build it | Yes, with LangGraph. MCP support is built in. |
LlamaIndex | Partial (loaders, plus managed connectors in LlamaCloud) | Partial (LlamaCloud SharePoint sync can read who has access to each file, and your app enforces it) | You build it | Yes |
Dify | Partial (data-source plugins) | Partial (set per knowledge base by workspace member, not by source ACLs) | Yes | Yes for both |
RAGFlow | Partial (Confluence, Notion, Google Drive, S3 and others) | Partial (a dataset is shared with "Only me" or "Team", not by source ACLs) | Yes | Yes for both |
Amazon Bedrock Knowledge Bases | Yes (SharePoint, Confluence, Google Drive, OneDrive, S3 and the web) | Partial (ACL-aware filtering on managed connectors) | Yes | Yes (AgentCore, with MCP) |
Azure AI Search | Partial (indexers, with SharePoint in preview) | Partial (document-level access, in preview) | Partial (optional answer synthesis with citations, off by default) | Agentic retrieval, and each knowledge base has an MCP endpoint |
Contextual AI | Yes (the on-demand plan allows one integration) | Yes (it syncs source permissions) | Yes (in-line attributions) | Yes, with an MCP server |
Glean | Yes (275+, per Glean) | Yes | Yes | Yes for both (hosted MCP) |
Onyx | Yes (50+ apps) | Yes, in the Enterprise Edition only | Yes | Yes for both |
PipesHub | Yes (50+) | Yes | Yes (block-level) | Yes for both (MCP server and client) |
Vectara | Partial (ingestion pipelines) | Partial (set per corpus; per-document control needs metadata) | Yes | Yes for both |
Sources for each row are listed at the end. Google Vertex AI Search belongs in the cloud group too, and it has a note below.
Frameworks you assemble
LlamaIndex. LlamaIndex is an MIT-licensed Python and TypeScript framework for getting documents into a shape an LLM can use. Its main strength is its excellent range of parsing and indexing options. Its paid LlamaCloud service adds hosted parsing, called LlamaParse, and managed connectors.
One of those is a SharePoint connector whose sync can read who has access to each file. LlamaIndex's December 2024 post says the permission check itself happens in your app. The current connector docs do not describe permission-filtered retrieval (checked 20 Nov 2026). Outside LlamaCloud, you build the user interface, logins, permission filtering and operations yourself.
LangChain. LangChain is the most widely used framework, with about 150,000 GitHub stars and an MIT licence. It includes LangGraph for building agents, and the paid LangSmith for tracing and evaluation. Its main strength is that it has an integration for almost everything, and its document loaders can read from hundreds of sources. Those loaders read a source once and do not keep it in sync, and permissions are up to you.
Haystack. Haystack is deepset's Apache 2.0 framework for building retrieval and agent pipelines in Python. Its main strength is clear, testable pipelines that suit production use. deepset also sells Haystack Enterprise Platform, priced by quote. It can run in the cloud, in a VPC, on your own servers, or air-gapped, which means fully cut off from the internet. As with other frameworks, connectors to business apps and permissions are your job.
RAG engines and builders
RAGFlow. RAGFlow is an Apache 2.0 engine with about 94,000 GitHub stars. It is known for deep document understanding. It parses scans, tables and complex PDFs, and it lets you inspect and adjust the chunks. It shows grounded citations, supports agentic retrieval and MCP, and syncs from sources including Confluence, Notion, Google Drive and S3. Version 1.0 shipped on 6 November 2026.
There are two things to know before you start. Its docs share a dataset with "Only me" or "Team", and they do not describe inheriting each source's per-user permissions (checked 20 Nov 2026). Its recommended starting server is 4 CPU cores, 16 GB RAM and 50 GB disk.
Dify. Dify is a popular open-source platform for building LLM apps, workflows and agents visually. It includes a "knowledge pipeline" for RAG. Its main strength is that it gives a mixed team a quick way to ship a working assistant. Since version 1.9, data-source plugins cover tools such as Notion, Confluence, Google Drive, Box and S3, and agents can use MCP servers.
Dify's licence is Apache 2.0 plus conditions. You need written permission to run it as a multi-tenant service, which is one service shared by several separate organisations. You also may not remove the Dify logo in the frontend. Knowledge access is set inside Dify, per knowledge base and by workspace member. Its docs do not describe inheriting permissions from source apps (checked 20 Nov 2026).
Managed and enterprise platforms
Vectara. Vectara is a RAG-as-a-service platform with an API for indexing, retrieval and grounded answers. It also created a well-known hallucination leaderboard, which tracks how often models state things their sources do not support. Its main strength is a hosted API that you can buy as SaaS (software the vendor hosts for you), in a VPC, or on-prem. Its pricing page lists SaaS from $100K a year, VPC from $250K and on-prem from $500K, with a 30-day trial.
Vectara's own docs say source-system permissions "do not automatically travel with each record". Access is set per corpus, which is Vectara's name for a collection of documents. For per-document control, you add metadata and filter on it yourself.
Contextual AI. Contextual AI is a managed platform for production RAG agents. It also sells its parts as separate APIs. These parts are parsing, reranking (re-sorting results by relevance), grounded generation and evaluation. Its docs say connectors sync permissions from the source and enforce them, so a user who cannot open a file cannot query it.
Pricing is pay-as-you-go with $25 of free credits, or custom Enterprise pricing with deployment in the customer's VPC. The on-demand plan allows only one data integration, and roles and single sign-on appear to be Enterprise features.
Glean. Glean is the best-known enterprise search and assistant product. It is a mature, polished product with a large connector catalogue and vendor support. It is sold per user per month, and connectors and a hosted MCP server are included in the seat price. Glean says it inherits and enforces source-system permissions.
Glean is closed source, and its prices are negotiated rather than published. There is no open-source self-hosted edition. Its Customer Hosted option runs inside your own GCP or AWS account, but Glean still deploys and manages it and calls it "not a traditional self-hosted model".
Onyx. Onyx is the most-starred open-source enterprise search project, with about 33,000 GitHub stars. Its Community Edition is MIT-licensed. It offers connectors to 50+ apps, agents, MCP and air-gappable self-hosting. Onyx announced it is joining Zoom and says it will remain open source.
According to Onyx's docs, syncing permissions from source apps is an Enterprise Edition feature. In the Community Edition, connector documents are visible either to everyone or only to the connector's owner.
PipesHub. PipesHub is our product. It is covered in its own section below so you can weigh it separately.
Cloud services. All three cloud services share one strength, which is that they live inside your existing cloud account and contracts. With each of them you build the user experience yourself, and each ties you to one cloud.
Amazon Bedrock Knowledge Bases now has a managed form with connectors for SharePoint, Confluence, Google Drive, OneDrive, S3 and the web. It offers ACL-aware filtering, which AWS describes as filtering and not as a security boundary. Your app must authenticate users and pass their identity.
Azure AI Search is a search engine you build on. SharePoint document-level access control is in preview. Microsoft notes that a revoked user may still see content until the next indexer run.
Google Vertex AI Search offers connectors such as Jira and Confluence. Its access control must be enabled when a data store is created.
How to choose a RAG platform: a checklist
Data sources. List the apps where your answers live. Check that each one has a connector that stays in sync and is not a one-off import. Check also that edits and deletions flow through.
Permissions. Ask how the platform knows who may see each document. Ask how fast a removed permission takes effect, and whether that feature is in the edition you will buy.
Accuracy evidence. Prefer dated public leaderboards, or tests you can rerun yourself. Note who runs the benchmark and which products are missing from it.
Deployment. Decide whether you need SaaS, your own cloud (VPC), your own servers, or an air-gapped setup. Check which LLMs you may use, including local models.
Cost model. Pricing may be per seat, per query, per page or by capacity. It may also be free software plus your own hosting and staff time. Model a year at your expected volume, and check that the licence covers the features you need.
How to test a RAG platform on your own data
A short pilot on real data tells you more than any vendor page, including this one.
Write 30 to 50 real questions. Give each one its source document or documents and the correct answer. Include some questions that need several sources. Include some that have no answer, to see whether the system admits it.
Measure retrieval. Check whether the expected source appears in the top results. This shows whether failures come from search or from the model.
Measure answer accuracy. Grade each answer as correct, partly correct or wrong, and check that each citation supports its sentence. An LLM can do the grading at scale, but have a person check a sample.
Test permission leakage. Ask the same questions as two test users with different access. It is a failure if restricted content reaches the wrong user in an answer, a snippet or a citation. Then remove access in the source app and time how long the content takes to disappear.
Record cost and speed. Note the cost and time per answer, and the effort it takes to run the system.
Where PipesHub fits, and where it doesn't
PipesHub is an Apache 2.0 open-source platform that connects to 50+ business apps. It gives people search and cited answers that are filtered by what each person may see. Access is checked when the query runs, against the source system's own permissions.
Your own agents get the same filtered context. They can reach it through an MCP server, through software development kits (SDKs) for Python, TypeScript and Go, or through a Retrieval API that returns cited passages.
Citations point to the exact block of the source. Retrieval uses a knowledge graph, which is a map of how pieces of information relate to each other. You can bring any LLM, including local ones. You can self-host with Docker Compose or Helm, or use PipesHub Cloud.
On accuracy, PipesHub ranked #1 on the public EnterpriseRAG-Bench leaderboard (Hugging Face) as of 14 Nov 2026, with 93.24 overall (95.0% answer correctness) among 43 entries. Onyx, also compared in this article, runs that leaderboard.
PipesHub also reports 92.7% on Google's FRAMES benchmark (824 multi-document questions), against 78.9% for the best of 18 standard RAG pipelines it tried. The PipesHub team ran that test itself and has published a write-up.
Choose PipesHub if you need permission-aware answers across company apps and want to self-host or keep data in your own cloud. It is a good fit when you want one context layer that serves both employees and AI agents.
Two administration features are part of the paid Enterprise Edition. They are audit logs, and single sign-on (SSO) with SCIM, which is a standard for creating and removing user accounts automatically.
Who should choose something else:
A team that wants to assemble its own pipeline in code should use a framework such as LlamaIndex, LangChain or Haystack.
A team that is already all-in on one cloud, or that wants a closed, fully managed product from a large vendor, may prefer that cloud's service or Glean.
A team that mainly parses a fixed set of scans and complex tables, or that builds many small LLM apps visually, may be better served by RAGFlow or Dify.
FAQ
What is the best RAG platform in 2026?
There is no single best RAG platform. For building your own pipeline, LlamaIndex, LangChain and Haystack lead. For an open-source builder, RAGFlow and Dify are the most popular. For permission-aware answers across company apps, look at enterprise platforms such as Glean, Onyx and PipesHub, or at your cloud provider's service if you are committed to AWS, Azure or Google Cloud.
What is the difference between a RAG framework and a RAG platform?
A framework is a code library that gives you building blocks, so you write and run the pipeline, interface and access rules yourself. A platform runs the pipeline for you and, in enterprise products, adds connectors to business apps, user permissions, citations and a ready-made interface.
What is the best open-source RAG platform?
It depends on the job. RAGFlow (Apache 2.0) is strong for parsing difficult documents, Dify is best for building apps visually but adds licence conditions, and Onyx (MIT) and PipesHub (Apache 2.0) are open-source enterprise platforms with business-app connectors. Check which features, such as permission sync, sit in each project's paid edition.
What is RAG as a service?
RAG as a service is a hosted API that handles ingestion, indexing, retrieval and grounded answers, so you send documents and questions and receive cited answers without running infrastructure. Vectara, Contextual AI and the managed cloud services, such as Amazon Bedrock Knowledge Bases, are examples.
What does permission-aware retrieval mean?
Permission-aware retrieval means the system only retrieves passages the person asking is allowed to see in the original app, such as a restricted Google Drive folder or private Slack channel. Without it, an AI assistant can quote documents to people who should never see them.
How do I evaluate a RAG platform on my own data?
Write 30 to 50 real questions with known source documents and answers, then measure whether the right source is retrieved, whether the answer and citations are correct, and whether restricted documents leak to users who lack access. Also time how long a removed permission takes to stop appearing.
Is there a public benchmark for enterprise RAG platforms?
Yes. EnterpriseRAG-Bench is a public benchmark with a leaderboard on Hugging Face, run by Onyx. It covers some products in this article but not all. Glean, Onyx, Dify, Vectara and Contextual AI do not appear on it as of 20 Nov 2026, so treat it as one input alongside your own tests.
Related reading
Sources
All accessed 20 November 2026 unless noted.
EnterpriseRAG-Bench leaderboard, Hugging Face: https://huggingface.co/spaces/onyx-dot-app/EnterpriseRAG-Bench-Leaderboard
PipesHub README (licence, connectors, benchmarks, Cloud): https://github.com/pipeshub-ai/pipeshub-ai
PipesHub docs, Retrieval API: https://docs.pipeshub.com/developer/retrieval-api ; MCP overview: https://docs.pipeshub.com/mcp/overview
PipesHub pricing (Community vs Enterprise): https://pipeshub.com/pricing (checked 14 Nov 2026)
LlamaIndex repository (MIT): https://github.com/run-llama/llama_index ; permission-aware SharePoint blog (Dec 2024): https://www.llamaindex.ai/blog/permissions-aware-content-retrieval-with-sharepoint-and-llamacloud ; LlamaCloud SharePoint connector: https://developers.llamaindex.ai/llamaparse/connectors/sharepoint/
LangChain repository (MIT): https://github.com/langchain-ai/langchain ; langchain-mcp-adapters (archived 17 Sep 2026, MCP support moved into
langchain[mcp]): https://github.com/langchain-ai/langchain-mcp-adaptersHaystack repository (Apache 2.0): https://github.com/deepset-ai/haystack ; Haystack Enterprise Platform: https://www.deepset.ai/blog/introducing-haystack-enterprise-platform ; MCPTool: https://docs.haystack.deepset.ai/docs/mcptool
RAGFlow README: https://github.com/infiniflow/ragflow ; data sources: https://ragflow.io/docs/add_data_source/data_source_configuration ; MCP: https://ragflow.io/docs/category/mcp ; sharing and permissions: https://ragflow.io/docs/guides/team/resource_sharing/resource_sharing_and_permissions
Dify licence: https://github.com/langgenius/dify/blob/main/LICENSE ; README: https://github.com/langgenius/dify ; data-source plugins: https://docs.dify.ai/en/develop-plugin/dev-guides-and-walkthroughs/datasource-plugin ; knowledge-base permissions: https://docs.dify.ai/en/use-dify/knowledge/manage-knowledge/introduction
Vectara pricing: https://www.vectara.com/pricing ; hallucination leaderboard: https://github.com/vectara/hallucination-leaderboard ; integrations and permissions: https://docs.vectara.com/docs/agents/integrations ; MCP server: https://github.com/vectara/vectara-mcp
Contextual AI pricing: https://contextual.ai/pricing ; connectors: https://docs.contextual.ai/connectors/overview ; MCP server: https://github.com/ContextualAI/contextual-mcp-server ; attributions: https://docs.contextual.ai/reference/glossary
Glean pricing: https://docs.glean.com/glean-core-suite-pricing ; connectors: https://www.glean.com/platform/connectors ; MCP: https://docs.glean.com/administration/platform/mcp/about ; deployment models: https://docs.glean.com/get-started/build/about-deployment
Onyx README: https://github.com/onyx-dot-app/onyx ; Enterprise Edition and permissions: https://docs.onyx.app/admins/permissions/understanding_permissions ; joining Zoom: https://onyx.app/blog/onyx-joins-zoom
Amazon Bedrock Knowledge Bases: https://aws.amazon.com/bedrock/knowledge-bases/ ; ACL filtering: https://docs.aws.amazon.com/bedrock/latest/userguide/kb-managed-acl.html
Azure AI Search document-level access: https://learn.microsoft.com/azure/search/search-document-level-access-overview ; pricing: https://azure.microsoft.com/en-us/pricing/details/search/ ; revocation lag: https://devblogs.microsoft.com/ise/sharepoint-doc-level-access/ ; answer synthesis and citations: https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-how-to-answer-synthesis ; MCP endpoint: https://learn.microsoft.com/en-us/azure/search/agentic-retrieval-how-to-retrieve
Google Vertex AI Search access control: https://docs.cloud.google.com/generative-ai-app-builder/docs/data-source-access-control
GitHub star counts from the GitHub API, 20 Nov 2026.








