Core Concepts
How knowledge works
How your content becomes searchable knowledge, and how an engine turns a question into a cited answer.
Knowledge Engines answer questions from your own content. You connect sources or write with Dock. Docklee processes the content in containers, and an engine searches the sources linked to it when your app calls Lynk.
Read this page before you design your engines. It explains why an engine answers only from what you link to it, and how each answer is checked against your sources.
- 01ContentConnected sources, and Dock writes that declare a container.
- 02ContainerChunks, extracts, and embeds, or builds OKF concepts.
- 03EngineSearches its linked sources and ranks what it finds.
- 04LynkReturns a cited answer, a stream, or the context itself.
Without and with Docklee
| Without Docklee | With Docklee |
|---|---|
| You build parsing, chunking, embeddings, and a vector store for each project. | Containers do the processing, and engines share them. |
| Each assistant keeps its own copy of the same documents. | A source connects once and links to any engine. |
| The model can cite things it never saw. | Citations are checked against the chunks the engine used, and invalid ones are removed. |
| Answers drift beyond your content. | The default mode answers only from your sources, and says so when the answer isn't there. |
| Keyword search and semantic search are separate systems. | An engine searches vector and OKF sources together and merges the results. |
What you send and what's stored
| You send | Docklee stores |
|---|---|
| A 40-page product manual, connected in Sources | Chunks with their page numbers and section titles, embedded in a vector container |
dock.capture("Refunds are available within 30 days.", type="knowledge") | A chunk in the container's API/SDK Feed source |
| A wiki connected to an OKF container | Concepts, one per topic, and the links between them |
On ingest
- Arrive. Content comes from a connected source, or from
Dockwith acontainer. See Capture knowledge. - Process. A vector container splits content into chunks, keeps page numbers and section titles, and embeds each chunk. An OKF container builds concepts and the links between them. See Containers.
- Wait for a link. Processed content belongs to a source. No engine reads it until you link that source to the engine. See Engines.
On query
- Search. The engine searches every source linked to it, by meaning in vector sources and by concepts and their links in OKF sources.
- Rank. The best results are re-ranked against your question, and the top chunks are kept.
- Answer. For
lynk.query, a model writes the answer from those chunks only, following the engine's answer mode. Forlynk.retrieve, the chunks themselves are returned. - Check. Each citation is checked against the chunks the engine used. Citations to anything else are removed, and confidence is adjusted to match. See Citations and confidence.
Where knowledge lives
| Knowledge | Lives in | Read by |
|---|---|---|
| A connected source | A vector or OKF container, chosen when you add it | Every engine it's linked to |
A Dock knowledge write | The container's API/SDK Feed source | Every engine the feed is linked to |
| A Precision Mode answer | Its container, prepared ahead of time | Every engine that reads the container |
Build with this flow
- Link only the sources an engine should answer from. Every linked source can show up in its answers.
- Keep one vector dimension per engine.
- Use
lynk.querywhen Docklee should write the answer, andlynk.retrievewhen your own model should. - Show citations in your UI so people can check an answer against the source.
- Treat a low confidence score as a reason to say "I'm not sure", not as an answer.
Next steps
You know how knowledge flows. Start by sending your own content.