Core Concepts
Containers
Where knowledge is stored and processed, and how to declare a vector or OKF container on a write.
Knowledge has to be processed before an engine can search it. A container is where that happens. Every piece of knowledge, from a connected source or from Dock, lands in a container, and engines read it from there.
There are two kinds of container.
| Vector | OKF | |
|---|---|---|
| Stores | Chunks and their embeddings | Concepts and the links between them |
| Searches by | Meaning | Keywords and links between concepts |
| Dimension | Yes, one per container | None |
| Good for | Long documents, help centers, mixed content | Catalogs, policies, and wikis with clear concepts |
Vector containers
Use a vector container when...
- Questions use different words than your content.
- The content is long or loosely structured.
A vector container splits content into chunks, extracts its structure, and embeds each chunk at a fixed dimension. The dimension is the length of each embedding.
| Supported dimensions |
|---|
256, 512, 768, 1024, 1536, 3072, 4096 |
Your plan sets which dimensions you can enable, and a write can only use a dimension that's enabled for your workspace. Higher dimensions capture more detail and use more storage.
If you leave the dimension out, Docklee uses your workspace default. An engine reads vector sources of one dimension only, so pick one per engine and keep to it.
OKF containers
Use an OKF container when...
- Your knowledge has clear concepts and relationships.
- Following links between ideas matters as much as matching words.
An OKF container stores knowledge in the Open Knowledge Format, an open specification where each concept is a Markdown file and links between files form a graph. Docklee builds the concepts and their links from your content. No embeddings are used, so an OKF container has no dimension. See What is OKF.
Declare a container on a write
A Dock write declares the container, never an engine:
from docklee import Dock
vector_dock = Dock(container="vector", dimension=1024)
okf_dock = Dock(container="okf")
vector_dock.capture("Refunds are available within 30 days.", type="knowledge")
okf_dock.capture("Every Atlas desk has a 5-year warranty.", type="knowledge")
vector_dock.flush_sync()
okf_dock.flush_sync()To set the container once for every Dock, use environment variables:
export DOCKLEE_CONTAINER="vector"
export DOCKLEE_DIMENSION="1024"| Field | Values | Notes |
|---|---|---|
container | vector or okf | Any other value is rejected. |
dimension | An enabled dimension | Vector only. Leave it out to use your workspace default. A dimension that isn't enabled is rejected. |
The API/SDK Feed source
The first write to a container creates an API/SDK Feed source for it in Sources. Later writes to the same container, and for vector the same dimension, go to that same feed.
No engine reads a feed until you link it. A write can never change an engine's answers before you decide it should.
Next steps
Your writes now land in a container. Link the feed to an engine to start reading from it.