Knowledge and grounding
Grounding is answering from your documents instead of from what a model happens to remember. In Olympusx AI a tenant keeps its documents in collections, inside vector databases it owns, filled by pipelines; agents and assistants name the collections they answer from, and every grounded answer lists the passages it used. This page explains how those parts work together.
The parts
Section titled “The parts”| Part | What it is |
|---|---|
| Vector database | The store. It belongs to one tenant, and holds that tenant’s collections and nothing else. Depending on how Olympusx AI was installed, each one is a database server of its own on a cluster you choose, or a database on the installation’s shared vector store. |
| Collection | A searchable set of documents in one vector database, embedded by one embedding model chosen when the collection is created. |
| Pipeline | What loads documents into a collection: uploads, or a URL, an S3 bucket, an NFS share or an SFTP server, on demand or on a schedule. |
| Embedding model | A model the tenant is granted and has deployed, which turns text into vectors. A collection cannot change its embedding model later, because every vector in it must come from the same one. |
| Reranker | An optional model that re-scores retrieved passages against the question, so the few passages the answer uses are the most relevant, not only the closest. |
| Graph | For each collection, a map of its documents, their sections, and the people, organisations and places they mention, built after the collection changes. |
See Knowledge.
From document to vectors
Section titled “From document to vectors”- A pipeline reads a document from its source.
- The document converter turns it into text, keeping its structure: headings, tables, lists.
- The text is cut into chunks of about 1,000 characters by default, with a small overlap so a sentence is not lost at a boundary.
- The embedding model turns each chunk into a vector, and the chunk and its vector are stored in the collection.
Code is indexed the same way, but cut at functions, methods and types rather than at a character count. See Code sources.
From question to grounded answer
Section titled “From question to grounded answer”- The question is embedded with the collection’s embedding model.
- The closest passages are retrieved from each collection the agent or assistant names: 6 by default, at most 20.
- Passages below the relevance floor are dropped. A question that matches nothing well is answered without passages rather than grounded on unrelated text. The floor is 0.5 by default; a platform administrator changes it in Settings, and can set one per embedding model.
- If a reranker is configured, it reorders the candidates and the best ones are kept.
- The answer model writes an answer from those passages, and cites them.
- The answer arrives with its sources: each passage, its document and its relevance, the cited ones first.
What it guarantees
Section titled “What it guarantees”- A tenant’s knowledge stays the tenant’s. Collections live in vector databases that belong to one tenant; another tenant’s agents cannot name them.
- Embedding happens on your models. The embedding model is a model the tenant deployed, on your clusters, unless a platform administrator granted an external embedding model.
- An answer says what it rests on. The sources come with every grounded answer, in the console, in assistant chat and over the API, so a reader can check the passage behind a claim.
- Personal data and secrets can be kept out. The tenant’s guardrail policy applies to the questions and answers, and code is scanned for secrets before it is stored.