Source formats
Input
text · file · URL
Documents can be stored as text, files, URLs, Markdown or code. Source type and optional URI remain part of the document record.
OpenLink / Knowledge
Knowledge, scoped.
Index workspace sources. Retrieve them within the workspace boundary.
Knowledge ServiceWorkspace scoped
Collection
product-context
workspace / atlas-web
architecture.mdready
api-contract.mdready
preview-notes.txtqueued
Search request
How is preview access scoped?
architecture.mdchunk 12
Project-bound browser sessions are authorized through the service boundary before access is relayed to the target runtime.
api-contract.mdchunk 04
The visible answer is hydrated from stored document content after the filtered vector lookup returns identifiers.
Illustrative result. It shows the stored-document citation pattern, not a live search.
Document lifecycle
From source material to scoped retrieval.
Each stage is persisted as document, chunk or job state. A ready result is the end of a traceable lifecycle, not an opaque upload.
01 · Source
Add a workspace document
Text, file, URL, Markdown and code sources are represented as documents in a collection.
0102 · Queue
Index as a tracked job
Documents move through queued, processing, ready or failed states instead of disappearing into a background task.
0203 · Chunk
Create deterministic chunks
The Knowledge Service splits indexable text into repeatable chunks and records their metadata.
0304 · Embed
Use the configured embedding provider
Embedding configuration is kept separate from chat-model configuration and checked against the backend dimension.
0405 · Retrieve
Search with server-generated scope
Queries are embedded, filtered by workspace and collection on the server, then hydrated with stored citation content.
05
Service boundary
One access boundary between product surfaces and vector storage.
Browser, Next BFF and Project VM calls reach the Knowledge Service. It owns authorization, workspace and collection filters, and the connection to metadata and vector storage.
Browser / BFF / VM
product calls
Knowledge Service
authorization + filters
ZOKERBASE + Zero
metadata + vectors
Metadata remains auditable
ZOKERBASE keeps ownership, source text, chunks and job state.
Vector access is not client access
Zero receives vector rows and non-sensitive identifiers through the service boundary, not directly from browser clients or Project VMs.
What is persisted
The technical primitives behind a knowledge workspace.
Lifecycle state
Jobs
queued → ready
Ingest, delete and reindex work is represented by jobs. Documents expose queued, processing, ready, failed and deletion states.
Runtime topology
Backend
standalone · distributed
Standalone is the default local topology. Distributed deployment and migration are explicit queued administrative workflows with validation and rollback behavior.