Answers that cite the document they came from

Knowledge bases

Playable below - no signup

A knowledge base is the material your assistants are allowed to read: the handbook, the product documentation, the tender answers you have already written, the sources your sector runs on. Scoped per group, counted so you can see what retrieval is working with, and stored in a database you can back up and export.

Try it below

Live demo

The knowledge layer, as an admin sees it

Seven sample bases across four categories, with the document counts that drive retrieval. This is the management view — the part you set up once and then mostly forget.

Sample knowledge bases. No documents are included and nothing is uploaded — the demo has no network access.

How a document becomes an answer

Four steps, and you can inspect every one

Ingest and chunk

A document is split into passages sized for retrieval rather than for reading. Chunking is token-aware, so a passage does not end mid-sentence and lose the context that made it findable.

PDF, DOCX, Markdown, HTML, plain text

Embed and index

Each passage gets a vector in your own PostgreSQL, using pgvector. The index lives with your data rather than in a search vendor’s cloud.

pgvector, in your own database

Retrieve and rerank

A question pulls candidate passages, and a cross-encoder reorders them by how well they actually answer it. Reranking runs on CPU, so this does not require a GPU box.

retrieve → rerank → top passages

Answer with citations

The answer names the passages it used. That is the difference between a system you can check and one you have to trust — and the reason a wrong answer is recoverable rather than mysterious.

every claim traceable to a source

Control

Who can read what, decided once

Scoped per group

A base is visible to the groups you choose. An assistant reading it can never surface a passage to someone who could not open the document themselves.

Grouped by category

Internal, product, commercial, legal. Categories keep the list navigable once an organisation has thirty bases rather than three.

Exports to a file

A base exports on demand in an open format. Handing your own material to an auditor should not require giving them an account.

Worth knowing

Where the boundary sits

FAQ

Worth knowing

What formats can we index?

PDF, Word, Markdown, HTML and plain text, plus pages pulled from the systems you connect. Scanned documents need to carry a text layer — there is no OCR step in the ingest path.

Where do the embeddings live?

In your own PostgreSQL, using pgvector. There is no separate vector database to run and no search vendor holding a copy of your material.

Can we keep a base away from a cloud model?

Yes. Point the assistants that read it at a local model and no passage reaches a provider. That configuration is why the embedding and reranking services are built to run on CPU.

How big can a base get?

Large enough that disk is the practical limit rather than the index. Plan for the storage the documents themselves need; the vectors are small next to the source material.

Index one folder and ask it something

The test worth running is a question you already know the answer to, on material you already trust. The citations tell you whether retrieval found the right passage.