Mastering AI Knowledge Bases with Dify: From Creation to Advanced Retrieval

Dify provides a visual RAG pipeline that lets users manage personal or team knowledge bases through an intuitive UI and quickly integrate them into AI applications.

Developers can upload internal documents, FAQs, specifications, etc., for structured processing and later LLM queries. Unlike static pre‑trained data, knowledge‑base content can be updated in real time, ensuring the LLM accesses the latest information and avoids outdated or missing answers.

When a user asks a question, the LLM first retrieves relevant chunks from the knowledge base based on keywords, supplying high‑relevance context that improves answer precision. This approach lets the LLM rely not only on its training data but also on dynamic documents and databases.

Core Advantages

Real‑time : Data in the knowledge base can be updated anytime, keeping the model’s context current.

Accuracy : Retrieval of relevant documents reduces hallucinations and yields higher‑quality answers.

Flexibility : Developers can customize the knowledge base content to match specific coverage needs.

Supported source formats include long‑text files (TXT, Markdown, DOCX, HTML, JSON, PDF), structured data (CSV, Excel), and online sources (web crawlers, Notion).

If your team already has an external knowledge base, you can connect it to Dify via the "Connect External Knowledge Base" feature.

Use Cases

For example, to build an AI customer‑service assistant, upload product documents to Dify and create a conversational app. Traditional development may take weeks, while Dify can complete the process in minutes.

Knowledge Base & Documents

In Dify, a knowledge base (Knowledge) is a collection of documents, each potentially split into multiple chunks. Documents can be uploaded by developers or operators, or synced from other data sources. If you already have a document repository, you can link it without re‑uploading.

Create a Knowledge Base

Steps:

Create the knowledge base by uploading local files, importing online data, or starting empty.

Specify a segmentation mode; the system will split long texts into chunks, previewable before finalizing.

Set indexing and retrieval options so that, when a query arrives, the system searches the indexed documents and returns highly relevant snippets for the LLM.

Supported file types for upload include TXT, Markdown, DOCX, HTML, JSON, PDF, CSV, Excel, and more.

Embedding

Embedding converts discrete variables (words, sentences, documents) into continuous vector representations, preserving semantic information and enabling efficient retrieval.

Embedding models are large‑language models specialized in vectorizing text.

For more details, see "Dify: Embedding Technology and Knowledge‑Base Design/Planning".

Metadata

Refer to the metadata documentation for managing knowledge‑base metadata.

1. Import Text Data

Click "Knowledge" → "Create Knowledge Base" in the top navigation, then upload local files or import online data.

Upload Local Files

Drag‑and‑drop or select files; batch upload is supported up to the limits of your subscription plan. Single file size limit: 15 MB . SaaS and Community editions differ in batch‑upload count, total documents, and vector storage limits.

Import Online Data

Dify supports importing from Notion or web pages. Once linked, you cannot add local files to the same knowledge base to avoid mixed data sources.

Later Import

If you lack documents now, create an empty knowledge base and add content later.

2. Specify Segmentation Mode

After uploading, the content must be segmented and cleaned. Segmentation splits long texts into manageable chunks, which improves retrieval efficiency and answer accuracy.

Two segmentation modes are available:

General Mode : Splits text into independent chunks based on a delimiter (default "\n"). Users can customize the delimiter with regex and set maximum length and overlap.

Parent‑Child Mode : Uses a two‑level structure—large parent chunks (paragraphs) provide context, while small child chunks (sentences) enable precise matching. The system first retrieves child chunks, then includes the corresponding parent chunk for full context.

Both modes allow custom preprocessing rules such as removing extra whitespace, URLs, and email addresses.

General Mode Settings

Segmentation Identifier : Default "\n"; customizable via regex.

Maximum Chunk Length : Default 500 tokens (max 4000).

Overlap Length : Recommended 10‑25% of chunk size.

Parent‑Child Mode Settings

Parent Chunk : Can be paragraph‑based or whole‑document; similar identifier and length settings.

Child Chunk : Default identifier splits by sentence; default max length 200 tokens.

After configuring, click "Preview Chunk" to view segmentation results.

3. Set Indexing Method & Retrieval Settings

Choose an indexing method (high‑quality or economical) and corresponding retrieval options.

Economic Indexing

Uses keyword‑based inverted indexing with 10 keywords per chunk; lower accuracy but no extra cost. Can be upgraded to high‑quality later.

High‑Quality Indexing

Employs embedding models to vectorize chunks, enabling semantic search. Supports three retrieval types: vector search, full‑text search, and hybrid search.

For more on embedding technology, see the referenced Dify documentation.

Q&A Mode (Community Edition only)

Generates Q&A pairs for each chunk, using a "Q‑to‑Q" matching strategy that improves handling of frequent or similar questions. Supports Chinese, English, and Japanese but cannot be used with economic indexing.

Retrieval Options

Vector Search : Converts queries and chunks into vectors and ranks by similarity.

Full‑Text Search : Keyword matching similar to traditional search engines.

Hybrid Search : Combines both methods; optional weight settings to prioritize semantic or keyword relevance.

Each retrieval type offers configurable TopK (number of returned chunks) and Score Threshold (minimum similarity). Rerank models can be enabled for an additional re‑ranking step, consuming extra tokens.

DIFY ETL: txt, markdown, md, pdf, html, htm, xlsx, xls, docx, csv</code><code>Unstructured ETL: txt, markdown, md, pdf, html, htm, xlsx, xls, docx, csv, eml, msg, pptx, ppt, xml, epub

For detailed ETL differences, refer to the official documentation.