Knowledge plugins for DeepSeek Harness
Bring documents, notes and search results into the conversation.
Knowledge plugins let the agent answer from material it was never trained on: your documents, a notes vault, a wiki, or live search results. Most work by indexing the content in advance and pulling the relevant passages into the conversation when a question comes up. The difference from memory plugins is the source: this is reference material you supply, not things the agent picked up along the way.
1 plugin
How to choose
Start with the formats and sources you actually have, since a plugin that reads Markdown well may do badly with scanned PDFs or tables. Find out where indexing happens: if documents or their embeddings are sent to an outside service, that matters for anything confidential. Prefer plugins that show which passages an answer drew on, so you can check it, and look at how the index is updated when files change.
Terms used on this page
- RAG
- Retrieval-augmented generation: fetching relevant passages first and giving them to the model along with the question.
- Embedding
- A numeric representation of a piece of text that lets similar passages be found by meaning, not exact wording.
- Chunk
- A section of a document, split to a size suitable for indexing and retrieval.
Questions and answers
Are my documents uploaded anywhere?
It depends on the plugin. Some index entirely on your machine; others send text to an embedding API or a hosted vector store. Retrieved passages are also sent to whichever model you use, as part of the prompt.
Why does the agent still get answers wrong with my documents connected?
Retrieval can miss the right passage, or return one that looks related but is not. The model can also misread what it was given. Check the cited passages, and if the right one keeps being missed, adjust how the content is split or how many results are returned.