The Short Answer
ibl.ai is a RAG platform for a company knowledge base that you run yourself: you own all the code and the data, the vector index and retrieval stay inside your cloud, on-premise or air-gapped network, and answers come from any LLM you choose. Glean, Onyx, Dify, LangChain and Amazon Bedrock Knowledge Bases also do RAG, on very different terms.
Those terms are what this page compares, from each vendor's own documentation: what kind of product each one is, where the index runs, who holds the source code, which models it can call, and whether the bill grows with headcount or with use.
What is a RAG platform for a company knowledge base?
A RAG platform for a company knowledge base connects your internal sources, such as SharePoint, Confluence and Google Drive, indexes them, retrieves the passages relevant to a question and has a language model answer from those passages with citations.
AWS describes RAG as "a technique that uses information from data sources to improve the relevancy and accuracy of generated responses" (Amazon Bedrock Knowledge Bases).
The answer lists for this query mix four different kinds of product under one name. Telling them apart is the first decision, because each one leaves a different part of the job to you:
| Kind of product | Examples | What you still have to supply |
|---|---|---|
| Managed knowledge app | Glean, Microsoft 365 Copilot, Dust | Little to build; the vendor runs and changes the software |
| Cloud RAG service | Amazon Bedrock Knowledge Bases | The application, sign-in and interface around the retrieval API |
| Open-source app you host | Onyx, Dify | Servers, upgrades and support, plus anything past the app's scope |
| Framework | LangChain | Everything: ingestion, storage, interface, access control, hosting |
| Platform you own | ibl.ai | The infrastructure it runs on; the source code is delivered to you |
Which RAG platforms do buyers compare for an internal knowledge base?
The RAG platforms named most often for an internal knowledge base are Glean, Onyx, Dify, LangChain, Amazon Bedrock Knowledge Bases and Microsoft 365 Copilot, with Cohere and Dust behind them.
Here is how each one answers the questions that decide a knowledge base, taken from its own documentation:
| Platform | Where it runs | Source code | Models |
|---|---|---|---|
| Glean | Glean's GCP tenant, or Customer Hosted in your GCP or AWS account, managed by Glean | Proprietary; no manual deploys or architecture changes | OpenAI, Azure OpenAI, Gemini, Claude and Amazon Nova, via Glean's key or yours |
| Onyx | Self-hosted: Docker, Kubernetes, Helm or Terraform; air-gappable | MIT (Community Edition); Enterprise Edition adds features | All major providers, plus Ollama, LiteLLM, vLLM |
| Dify | Self-hosted (Docker Compose), Dify Cloud, or Dify Enterprise | Apache 2.0 with added conditions | Hundreds of providers; any OpenAI-compatible model |
| LangChain | Wherever you deploy the code you write | MIT framework; LangSmith is a separate commercial product | Interchangeable integrations |
| Amazon Bedrock Knowledge Bases | AWS, managed or with your own vector store | Proprietary AWS service | Service-managed by default, or your choice on Bedrock |
| Microsoft 365 Copilot | Microsoft's cloud, as a Microsoft 365 add-on | Proprietary | Microsoft-hosted by default; admins can enable Anthropic models |
| Dust | Hosted, US or EU data residency; single-tenant on Enterprise | Not a self-hosting offer on its pricing page | 20+ frontier models, including GPT, Claude, Gemini and Mistral |
| Cohere | Your VPC, or on-premise air-gapped behind your firewall | Proprietary | Cohere's model family; ask which others North supports |
| ibl.ai | Your cloud, on-premise or fully air-gapped | Delivered to you; you own all the code and the data | Model-agnostic: any LLM, switch anytime |
Sources: Glean deployment models, Glean LLM availability, Onyx on GitHub, Dify on GitHub, LangChain on GitHub, Amazon Bedrock Knowledge Bases, Microsoft 365 Copilot, Microsoft 365 model choice, Dust pricing, Cohere private deployments.
Which RAG platforms can we self-host?
Of the RAG platforms on the list, Onyx, Dify and LangChain can be self-hosted from public source code, and ibl.ai is self-hosted with its source code delivered to you.
Glean and Cohere run inside your own cloud account or network, but as proprietary software the vendor builds and licenses.
Onyx is the closest open-source match to a knowledge-base app. Its Community Edition is "available freely under the MIT license" and covers RAG, chat, agents and actions; it indexes "50+ applications" and offers an "air-gappable, self-hosted deployment" (Onyx). An Enterprise Edition adds features for larger organizations.
Dify is an open-source LLM app builder with a RAG pipeline, and its Docker Compose install needs 2 CPU cores and 4 GiB of RAM. Its license is Apache 2.0 with conditions: "you may not use the Dify source code to operate a multi-tenant environment" without written permission, and you may not remove the logo from the console (Dify license).
LangChain is "a framework for building agents and LLM-powered applications," MIT-licensed (LangChain). It self-hosts because you write and deploy the application yourself; ingestion, access control and the interface are your team's to build.
Glean offers Customer Hosted, where "Glean deploys its tenant as a managed service in isolation within your own cloud environment" on GCP or AWS. Glean remains "responsible for infrastructure and application management, including upgrades, patching, and monitoring" (Glean).
Running in your account answers where the data sits. It does not answer who can change the software, fork it, or keep it running if the vendor relationship ends.
Our Glean alternative guide and Onyx alternative guide go deeper on each.
How much does a RAG knowledge base cost per month?
A RAG knowledge base costs either a fixed price per employee or a usage price per gigabyte and query, and the two shapes diverge fast as headcount grows.
Microsoft 365 Copilot lists $30 per user per month paid yearly (Microsoft); Amazon Bedrock Managed Knowledge Bases charge $5.00 per GB of raw data per month and $1.00 per 1,000 Retrieve calls (AWS pricing).
Per-seat pricing is the wrong shape for a knowledge base, because the bill counts people who might ask a question, not questions asked. The table below applies each published rate; the usage rows assume 50 GB of source documents and 20 retrievals per employee per month.
| Pricing shape | 1,000 employees / month | 5,000 employees / month |
|---|---|---|
| Microsoft 365 Copilot ($30/user, plus the required Microsoft 365 plan) | $30,000 | $150,000 |
| Dust Pro (€24/seat yearly; Business plan covers up to 100 people) | Enterprise quote | Enterprise quote |
| Bedrock Managed Knowledge Base (storage + retrieval, before model tokens) | $270 | $350 |
| ibl.ai, self-hosted (your infrastructure + model tokens, no per-seat fee) | Grows with use | Grows with use |
The Bedrock rows are 50 GB × $5.00 = $250 of storage, plus 20,000 or 100,000 retrievals at $1.00 per 1,000. Model tokens come on top and depend on the model and answer length, which is exactly why a usage bill tracks the work and a seat bill does not.
The same arithmetic holds for any self-hosted option: Onyx, Dify and ibl.ai carry no per-seat fee in the software, so the cost is servers plus tokens. For a fuller per-seat comparison at scale, see Enterprise AI with no per-seat pricing.
Which RAG platform should a company choose for its knowledge base?
The right RAG platform for a company knowledge base depends on how much of the stack the company wants to own.
Per Glean's documentation, even its Customer Hosted tenant does not support "altering any part of the Glean architecture," so ownership is the axis that splits the list most cleanly.
- Choose a managed app (Glean, Microsoft 365 Copilot, Dust) when speed matters more than control and a per-user bill is acceptable at your headcount.
- Choose Amazon Bedrock Knowledge Bases when you are standardized on AWS and have a team to build the application around a retrieval API.
- Choose Onyx or Dify for a self-hosted pilot from public source code, and read the license terms first, Dify's multi-tenant condition in particular.
- Choose LangChain when retrieval is one component of a product your engineers are building anyway.
- Choose ibl.ai when the knowledge base has to run inside your perimeter, on any model, with the source code in your hands and no per-seat bill.
Our Fortune 500 knowledge base guide covers the control pattern at 50,000 employees, and the open-source AI search engine guide compares Onyx, Haystack, LlamaIndex and txtai.
How does ibl.ai run a RAG knowledge base on your own infrastructure?
ibl.ai runs a RAG knowledge base as part of Agentic OS: connectors bring in your systems, a knowledge graph and vector index provide policy-aware retrieval, and agents answer while citing the source they used. 1.6M+ users across 400+ organizations run the platform, including NVIDIA, MIT, and Syracuse University.
You receive source access to the connectors, policy engine and agent interfaces, so the index, the retrieval logic and the logs sit on infrastructure you control.
The platform is model-agnostic: route sensitive questions to a local model and others to Claude, GPT or Gemini, and change the routing as models improve.
Because there is no per-seat pricing, the bill does not grow when you add the next department. ibl.ai is family-owned and operated from New York, NY, and the same stack runs in your cloud, on-premise or fully air-gapped, which our air-gapped platform comparison covers.
What should a buyer ask a RAG platform vendor before signing?
A buyer comparing RAG platforms for a company knowledge base should ask five questions before signing. Each one maps to a column in the tables above:
- Where does the vector index live, and who can reach it? Glean's Customer Hosted docs say Glean employees can view non-PII system logs for debugging, and that blocking Glean's access entirely means Glean "can't investigate issues" (Glean).
- Do we receive the source code, and under what license? MIT, Apache 2.0 with conditions, and proprietary are three different answers.
- Can we use any model, including one we host? Onyx names Ollama, LiteLLM and vLLM; Dify supports any OpenAI-compatible model.
- Does the bill scale with employees or with queries? At 5,000 employees, $30 per seat is $150,000 a month before a single question.
- What happens to the index and the connectors if we leave? A platform you own has no exit to negotiate.
To size it for your own sources and headcount, talk to the ibl.ai team.