Skip to main content

Knowledge Bases & Documents

A knowledge base is a collection of documents, an Agent canvas, and retrieval configuration. The knowledge base list page shows all knowledge bases and their document counts:

Knowledge base list

Create a knowledge base​

Click "New knowledge base", enter a name and description, and optionally specify an Embedding model (leave empty to use the team default):

Knowledge base edit drawer

  • The chunking strategy and chunk size cannot be changed after creation, so choose them based on your document types;
  • Changing the Embedding model clears existing vectors and re-indexes all documents;
  • A knowledge base supports two visibility levels: "private / team-shared".

Upload documents​

Open a knowledge base's detail page and drag & drop, paste, or select multiple files to upload. Documents enter an asynchronous pipeline: parse → chunk → embed → index, with status shown in real time in the list:

Document list

Batch selection, enable/disable documents, re-parse, metadata editing, and chunk preview are supported.

Curated Q&A​

Maintain human-curated standard Q&A (question-answer pairs) for a knowledge base. When a user question closely matches a curated question, the preset answer is returned directly, skipping retrieval and model generation — keeping high-frequency answers stable and consistent, and suiting tightly-constrained scenarios such as customer-service scripts.

Retrieval testing​

Each knowledge base has a built-in retrieval testing page where you can compare the balance of keyword and semantic weights and inspect the hits at each stage (dense / lexical / fused). It is the primary tool for tuning retrieval quality.

Citation tracing​

Every citation under an answer is clickable: clicking opens the source document viewer on the right, positioned at the cited passage:

Document viewer — original PDF rendering

  • Original mode: PDF.js renders the original PDF with page navigation, highlighting the cited passage right on the page;
  • Parsed mode: the parsed output is paginated, with the hit passage highlighted and scrolled into view;
  • Side by side: on desktop the document panel sits next to the conversation; on mobile it takes the full screen.

Scanned documents where the passage cannot be located degrade gracefully to page-level annotation.