retrievr

How it searches

RAG without running a vector database

The honest answer is that vectors do not go away. Comparing meaning is how a paraphrase gets found, and there is no way to do it without something that stores and compares vectors. What you can skip is operating that thing: the embedding job, the index, the rebuild after a model changes, and the tuning that decides how big a chunk should be.

So this page is about the trade rather than a slogan. Below is what the retrieval actually does, what leaves your side of the wire, and what you give up by not owning it.

Three ways of looking, and a rule for combining them

A single retrieval method fails in its own direction, and the failures are not small. A vector search drifts to whatever is nearest in meaning, so a question with an exact answer in the text can return a passage that merely discusses the subject. A keyword search cannot see a paraphrase at all: ask about a notice period and a document that says “either party may terminate on thirty days’ notice” is invisible to it unless the words match.

Retrieval here runs three legs over your documents and fuses them: a vector leg for meaning, a lexical leg for the words that were used, and a phrase leg where the terms must be adjacent — which is what lets an exact quotation win over a passage that covers the same ground.

The fusion is on rank, not on score, and that is the part most implementations get wrong. A cosine distance and a text-search rank are not comparable quantities, and adding them together produces a number that means nothing. Each leg proposes its candidates best-first, and their positions are combined — so a passage that two legs agree on rises, and one that a single leg loves cannot dominate.

What you never have to do

No embedding job to run, no batch to monitor, no worker to keep alive. No index to create, and nothing to rebuild when an embedding model is replaced — that is a migration on our side, and your documents are re-embedded as part of it.

No chunk-size decision either. That one is a real trade made on your behalf: text is packed to about 400 tokens and cut where the document changes section, never through the middle of a clause. You do not get to set the number. How a document is read is the long version of that decision.

And no database to size. Adding a thousand documents is an upload, not a capacity planning exercise.

What it costs you, beyond the cent per query

Control, in three specific ways, and it is fairer to say so than to leave it to be discovered.

You cannot tune the retriever. The number of passages, and which documents a search covers, are yours to set — the fusion, the passage size and the embedding model are not. A team that wants to weigh a leg differently, or to try a different embedding model against its own corpus, wants a library it can run rather than a service.

You cannot audit the index. You can audit every result, which is what a citation is for, but not the structure it came out of. If a compliance process requires the ability to inspect the store itself, this is the wrong shape of product.

And the documents live here rather than in your warehouse. They are yours, they are searchable only for your account, and they can be deleted for good — but they are not a dataset you can join against.

What you do instead

Two lines, and no infrastructure:

Fetch https://retrievr.io/install and follow it.

Then your agent asks in plain language and gets passages back with their source, author, section, character range and licence. A search returns about three passages of roughly 400 tokens by default, so the cost of an answer is what the model reads rather than what the document weighs.

If you are building the client yourself, the protocol reference is the raw shape: POST /retrieve, then poll the job. An MCP-speaking harness needs neither call — the PDF server page covers what comes back for a document.

When the answer is that you should run one

If your corpus is measured in millions of documents, if the retrieval feeds a product whose ranking is the product, or if a lawyer needs the index to exist somewhere you can point at — build your own. The pieces are well documented and the work is real; it is simply a different job from answering a question about a contract.

The reason to hand it over is that the interesting part of your application is almost never the retriever, and operating one is a standing cost that grows quietly.