Agent knowledge bases are in beta.
An agent knowledge base processes your content for you: a reasoning model reads and organizes each document, and an embedding model makes it searchable. Your part is small: choose what goes in, set how deep it searches, and check what comes back.
At the default depth, a search matches by meaning and keywords. Set a deeper retrieval depth, and the knowledge base investigates a question the way an agent would: following related subjects across documents, or breaking the question into parts and reading whole sections.
Besides your apps, coding agents on your own machine can read its documents through the Dify CLI, browsing them like files in a folder.
-
To weigh it against the classic knowledge base, see which kind to create.
-
To move a classic knowledge base over, see how to upgrade.
Set Up#
-
Create an agent knowledge base: Connect websites, online documents, or online drives that sync on your schedule, or upload files.
-
Review indexed documents: Check how documents were split and summarized, and fix what didn’t come through well.
-
Set up image retrieval: Choose models that let retrieval work with images.
Test and Improve#
-
Test and tune retrieval: Try what your users will ask, and save the ones that must keep passing after every change as golden questions.
-
Adjust retrieval settings: Choose the models and the retrieval depth, and set how many chunks come back and how closely they must match.
Use and Maintain#
-
Connect to apps, agents, and code: Add it to workflows and new Agents, call it from your own code, or let coding agents read it.
-
Maintain an agent knowledge base: Monitor the knowledge base once apps depend on it, fix what goes wrong, and keep its content up to date.