RAG as a Service Platforms Compared: Which Type Suits Your Organisation?

Asking for the best RAG platform is a bit like asking for the best vehicle: it depends what you are carrying. RAG as a service platforms come in three types: managed pipelines inside a cloud account, a hosted vector database combined with an open-source framework, and knowledge-base marketplaces where a base answers questions minutes after upload. This guide compares them on set-up time, pricing model, citations, API and MCP access and data control under UK GDPR, then gives you a decision grid to shortlist the right type.

Three types of RAG as a service

If the term itself is new, read RAG as a service: what it is and when to use it first. Here we take the next step: assuming you want to outsource some or all of the retrieval pipeline, which type of offering should you be evaluating?

Managed pipelines from the cloud providers

Amazon Bedrock Knowledge Bases, Google Cloud's Vertex AI Search and Azure AI Search are familiar names in this type. You connect a data source such as a storage bucket or SharePoint, pick an embedding model and an index, and the provider runs ingestion and retrieval. Your application calls the API and decides how answers and sources are displayed.

The appeal for larger organisations is that it sits within an account the IT and security teams already govern: identity management, audit logs, UK or EU regions, existing data processing agreements. The trade-off is that configuration and the user-facing application remain your responsibility.

Vector database plus framework

Here only the storage layer is a service. A hosted vector database (Pinecone, Weaviate Cloud, Qdrant Cloud, or Postgres with pgvector) holds the embeddings, while a framework such as LangChain or LlamaIndex glues together parsing, chunking, retrieval and the model call. You write and maintain the code, and you make every design choice, from chunk size to reranking to how citations are formatted. Our explainer on vector databases, embeddings and indexes covers the vocabulary.

Knowledge-base marketplaces

The third type hides the pipeline behind a knowledge base you can query straight away. Upload documents, and the platform extracts, chunks and indexes them; the base answers on the web, by API and frequently through an MCP server, with cited passages. A marketplace also offers a catalogue of bases built by others, usually priced per question. Kopik belongs to this type.

The comparison table

We compare types rather than named vendors: products evolve every quarter, the structural trade-offs much less. Prices are deliberately absent, as they depend on volume, region and contract. Always check the provider's current pricing page.

Criterion by criterion

Set-up time

Measure the time to a correct, sourced answer for a real colleague, not the time to a demo. Framework stacks demo quickly and then absorb weeks of chunking and evaluation work. Cloud pipelines shorten ingestion but still need an interface. Marketplaces skip the plumbing if their defaults suit your documents, which is worth testing on 20-30 real questions before deciding.

Pricing model

Ask every supplier the same question: what does one answered question cost, all in, including generation? Cloud pipelines meter several services; framework stacks hide the largest cost in engineering time; marketplaces tend to bill per question. Also ask how the bill behaves if your corpus doubles or usage spikes during a busy month.

Citations

For HR, legal or compliance content, citations are not a nice extra. Check that each claim points to a numbered passage, that you can read the raw passage text, and that the system admits when the documents do not contain the answer. Without that admission, a fluent answer and an invented one look identical.

API and MCP access

An API is standard. MCP support is the differentiator in 2026: the Model Context Protocol lets assistants such as Claude, Cursor or ChatGPT call a knowledge base as a tool. A native server means one line of configuration; otherwise someone builds and hosts a wrapper. See the MCP explainer for UK teams.

Data control under UK GDPR

If documents contain personal data, you remain the controller and the platform is your processor, so you need a written contract covering the Article 28 terms, clarity on where data is stored and any international transfers, and an answer on whether content trains models. The Information Commissioner's Office publishes guidance on AI and data protection worth reading before a data protection impact assessment. Cloud pipelines inherit your existing agreements; marketplaces must explain how private bases are isolated.

A decision grid

1. Retrieval is your product's core feature, or your data is unusual? Vector database plus framework.
2. Everything must stay inside one cloud tenancy and UK or EU region? That provider's managed pipeline.
3. You need sourced answers this month from existing documents, for staff or AI agents? A knowledge-base platform.
4. You need expert material you do not own, such as HMRC or HSE guidance? Check a marketplace catalogue before building.
5. You want to charge for access to your expertise? Only a marketplace with creator payouts does this.

Mixing types is normal: a product team might build a custom stack for the customer-facing search while operations staff query a hosted private base from their assistant.

Where Kopik fits

Kopik is a knowledge-base marketplace. Creating a base is free: upload PDF, Word, text or Markdown files and Kopik extracts, chunks and indexes them for hybrid search, combining full-text search with semantic expansion of the question's keywords. Answers come with cited passages, or you can request raw passages only.

• Private bases are accessible only to you and your API keys.
• Public bases sit in the catalogue, priced per question by their creator, who keeps 70%. Examples include UK VAT for businesses from HMRC notices.
• Access via the website, a REST API, or the MCP server at https://kopik.io/api/mcp.
• Pricing: prepaid credits per question, or a €12 monthly chat subscription with a usage gauge.

Kopik does not run in your own cloud account and does not let you swap embedding models. If you need either, look at the first two types.

Questions to put to any RAG supplier

Whichever type you shortlist, send every supplier the same short questionnaire. The answers are easy to compare and quickly reveal gaps that a sales demo will not show.

• What does one answered question cost, all in, including generation?
• Can I retrieve the raw passages behind every answer, not only the written response?
• What does the system do when the documents do not contain the answer?
• Is there a native MCP server, and which transport does it use?
• Where is data stored, are there transfers outside the UK, and is content used for training?
• How do I delete a document, and how quickly does the index reflect it?
• Can I export everything and leave without friction?