Comparison
A ChatGPT alternative for when documents cannot leave the company
This is not a contest about which is smarter. It is a choice between two architectures: one sends your text to a vendor, the other keeps it on your server. Here is when the second one wins — and when it does not.
ChatGPT is an excellent tool and for a great many tasks it remains the better choice. This page does not try to talk you out of that: it tries to isolate the narrow case where a self-hosted system is genuinely more suitable.
The issue is not trust, it is structure
Serious hosted AI vendors state that they do not use business customers' data to train their models, and there is no reason to doubt it. The point is different: even in the best case, that text has still left your organisation and come under someone else's control.
From then on your security posture includes a party you do not administer. Contractual terms can change, so can the set of sub-processors, and each time it falls to you to reassess. That is not an accusation against anyone: it is the structure of the relationship.
For most office tasks that structure is perfectly acceptable and convenience wins. For some categories of document it is not — and the distinction is worth drawing in advance, rather than after somebody has pasted the wrong thing.
When you do NOT need a self-hosted system
Worth stating plainly, because standing up a server for the wrong use case is simply wasted work.
- If the material is public or otherwise not confidential, the convenience of a hosted service wins without argument.
- If the task is open-ended and generative — writing, brainstorming, translating, producing code from scratch — a frontier model is stronger, visibly so.
- If nobody in the organisation can administer a server, a managed service is the honest choice.
- If there are few enough documents to read in an afternoon, there is nothing to index.
The differences that actually matter
Reduced to the six that move something in a business decision. The rest are details.
| Criterion | Hosted service | Self-hosted RAG |
|---|---|---|
| Where the document ends up | On the vendor's servers | On your server, and it stays there |
| General reasoning | State of the art | Good, not state of the art |
| Search across your archive | Over what you upload, within plan limits | Over the whole indexed archive, no plan limits |
| Cost structure | Subscription and usage, forever | Hardware once, then electricity |
| Works without a network | No | Yes, after the first run |
| Auditability of the code | Documentation and certifications | Public source, readable line by line |
A comparison between two ways of delivering the service, not between two named products. Hosted services' contractual terms change over time: check the vendor's current documentation before deciding.
When it is the right choice
- When the material is covered by professional privilege, a contractual confidentiality obligation or sector regulation, and any transfer to a third party has to be justified in writing.
- When the archive holds personal data, especially special categories such as health data, where an impact assessment is required anyway and is easier over a closed perimeter.
- When the network is isolated by choice or necessity, and an external service is simply not reachable.
- When internal query volume is high and continuous, and metered pricing becomes a budget line that grows on its own.
- When you need to show an auditor what the software does by reading the code, rather than producing a certificate.
In many organisations the right answer is not to choose: use the hosted service for general work and an internal system for confidential archives, with a written rule saying which material goes where.
What the Community edition concretely gives you
The Community edition is open-source under AGPL-3.0 and usable in production without a commercial agreement. It is not a demo build with features switched off.
- A single executable to run: it starts the vector database and the inference engine itself, with no Docker and no containers to orchestrate.
- Upload of PDF, DOCX, XLSX, HTML, TXT, Markdown and CSV, with automatic OCR in Italian and English on scanned pages.
- Multilingual semantic search across 100+ languages, so an Italian question finds the passage inside an English document.
- Streamed answers with sources cited, and conversations that keep context across follow-up questions.
- Authentication and three roles, separating who may only read from who may upload or delete.
- A consistent daily backup and restore over the running install, with archives kept as local files.
If you then need structured extraction, hybrid search, advanced roles, audit and a support contract, that is the Pro edition — but the right way to get there is to start with Community on real documents.
Frequently asked questions
- Is there a ChatGPT alternative that runs inside the company?
- Yes. A self-hosted RAG system puts the language model, the index and the documents on the same company server. The I3K RAG Enterprise Community engine is open-source under AGPL-3.0, installs as a single executable, and works without an internet connection after the first run.
- Is data entered into ChatGPT used to train the models?
- For business plans, vendors state that customer content is not used for training, and the terms are published. The point that matters for a corporate assessment is still a different one: regardless of training, that content was transmitted to an external party, and that has to be recorded in your processing register and supported by a legal basis.
- Is a local model much worse than ChatGPT?
- On open-ended reasoning and creativity, yes — the gap is real and visible. On the specific task of answering questions about documents you supplied, the gap narrows considerably, because there quality depends mostly on retrieving the right passages rather than on the size of the model.
- How does the cost compare with a subscription?
- The structure differs: no per-user fee, but an upfront hardware investment. The comparison only means anything on your own numbers — how many people, how many questions a day, which documents — and should span at least two or three years, because that is where a one-off spend and a recurring fee cross over.
- Can we use both?
- That is the most common arrangement and often the most sensible. The hosted service covers general work while the internal system covers confidential archives. What it needs is a written, known rule about which material may go where, otherwise the separation exists only on paper.
Ready to run RAG on your own infrastructure?
Start with the open-source Community edition, or talk to us about Pro with structured extraction, SSO, audit log and SLA.