RAG applications
Retrieval apps without the plumbing.
Ship apps that answer from your own content and show their work. Put documents in a collection and every answer comes back with the passages it used. Underneath, Akumi ingests them into a knowledge graph walled off per workspace, so a request can only ever reach your own. No vector store to run, and nothing leaves the EU.
collections · one graph · EU-resident
Added to the prompt as context, before the model answers.
RAG is easy to demo and a project to run.
Behind a working demo sits a vector database to operate, an embedding pipeline to maintain, a chunking strategy to tune, and often a US embeddings API that ships your documents abroad. Then you still have to keep one team's content out of another's answers and prove where any of it was processed.
What you get
Retrieval as a managed service.
The parts of a retrieval stack that take a quarter to build and a year to maintain are the parts you do not run here.
No pipeline to run
Paste or upload PDF, text or markdown into a collection. The platform extracts, chunks and writes it into the knowledge graph in the background. No vector store, no embedding pipeline, no chunking strategy to tune.
Walled off by construction
Every collection is its own partition of the graph, namespaced to your organization and workspace before the query is built. A request can only ever reach your own partitions, never another tenant's, and the graph stays in the EU.
Grounded and cited
Retrieval runs on its own whenever the workspace has something to consult, blending its facts, documents and collections into one ranked context. The response carries the sources it used, so you can show where an answer came from.
What it looks like
One request, start to finish.
An insurer's support assistant answers a policyholder from the team's own documents. You send the question. You get the answer, and the passages it stands on.
- 01SendYou send a questionOne OpenAI-compatible call. The user field is what scopes personal memory, and the workspace and its collections come from the API key.
- 02ConsultAkumi consults the graphRetrieval blends that end-user's memory, the workspace and every collection in it, then ranks what comes back. Nothing to orchestrate.
- 03CiteYou get the receiptsThe answer arrives with a sources array, and the request is recorded with the model, provider and region that served it.
No rag parameter, no retrieval call, no collection to name.
Which service does what
Retrieval, and everything around it.
A real retrieval app is more than a similarity lookup. The services it needs sit behind the same endpoint, each metered on its own.
- RecallPer-user facts and documents, promotable into a collection.
- Inference APIOpenAI-compatible generation on EU-resident models.
- FirewallPseudonymizes personal data in retrieved context.
- ObservabilitySee whether time went to retrieval or the model.
- CacheReuse answers for repeated questions.
- Audit TrailWhich request read what, on the record.
FAQ
Questions people ask about retrieval.
- Do we run a vector database?
- No. Put documents in a collection and the platform extracts, chunks and indexes them in the background. There is no store to operate and no pipeline to maintain.
- Can one team's content reach another team's answers?
- No. Every collection is its own partition, namespaced to your organization and workspace before the query is built.
- Do we have to call retrieval ourselves?
- No. Retrieval runs whenever the workspace has something to consult. There is no rag parameter and no collection to name in the request.
- How do we show where an answer came from?
- The response carries a sources array with the documents and passages it used, and their relevance.
- Where are the documents processed?
- In the EU. The embedding models and the graph are EU-resident, and every request records the region that served it.
Start here
Ship retrieval, skip the infrastructure.
Create a key, put your documents in a collection, and your app answers from them.