Track 6 · Advanced

6.1.4

Dify

Open-source LLM app builder — upload papers, get a grounded chatbot

7 lessons 2026-08-06 AI-generated

1Overview

Built by LangGenius, open-sourced in 2023 — the pitch is one stack instead of stitching several tools together to get a chatbot talking to your own documents.

Dify is an open-source platform for building LLM-powered applications — chatbots, document Q&A, agent pipelines — through a visual interface. It handles the whole stack: prompt management, RAG with a built-in vector store, tool calling, conversation memory, and a publishable chat UI. Upload twenty PDFs, let Dify chunk and embed them, and ask a chatbot to synthesise findings across the corpus with citations. 50+ model providers (OpenAI, Anthropic, Groq, local Ollama) swap behind one UI. → Where n8n excels at API-level plumbing, Dify is LLM-first: the knowledge base is the product. Apache 2.0 and self-hostable; the cloud free tier is 200 messages a day.

Dify is the LLM-first member of the automation family: a visual builder for chatbots, document Q&A and agent pipelines, with RAG and a vector store built in. Below: what it is best at, what to watch for, and its full lesson set.

1.1Before you start

Automation tools— Reading Automation tools introduces LLM app automation concepts, preparing you to understand Dify's capabilities.

1.2After this chapter you can

→Build a grounded Q&A chatbot over your own PDF, no code

→Assemble a knowledge base from many documents and cite its answers

→Swap the model provider without rebuilding the app

→Chain a multi-step agent workflow on the visual canvas

1.3Best for

Standing up a document Q&A chatbot without picking a vector database or writing retrieval code yourself — the RAG pipeline is already wired in.

1.4Watch out

The cloud free tier (200 messages/day) is exhausted quickly in a classroom. Retrieval quality depends on the chunking strategy — defaults need tuning for dense scientific PDFs — and self-hosting needs Docker and ≥4 GB RAM. The course instance (dify.32dots.de) is currently paused.

1.5Free vs paid

Cloud free tier: 200 LLM messages/day, unlimited RAG queries. Professional $59/mo. Self-hosted: fully free under Apache 2.0.

2Lessons 7

2.1 Add a custom Azure AI Search tool to your Dify workspace

A custom tool that lets a Dify app call Azure Cognitive Search via its REST API.

You will be able to invoke Azure AI Search from a Dify workflow and see search results returned as tool output.

Open your Dify workspace and navigate to Integrations > Tools.
Click “Add tool” and choose the Custom Endpoint option.
Paste the OpenAPI schema shown in the Microsoft AI Search guide (including title, version, server URL, path /indexes/wukong‑doc1/docs, query parameters api-version and search).
Save the tool and give it a name like “AzureSearch”.
Create a new workflow, add a User Input node followed by a Tool node that selects the AzureSearch tool you just created.
Configure the Tool node to map the user’s input to the “search” query parameter and set a fixed api‑version value.

You'll see When you run the workflow with a search term, the tool node returns a JSON response containing an array of matching documents from Azure AI Search.
Takeaway Custom OpenAPI schemas let Dify extend LLM capabilities with any external service that provides a REST API.

2.2 Run Dify locally with Docker Compose

A self‑hosted instance of the open‑source Dify platform running in Docker containers.

You will have a running Dify service accessible at http://localhost:3000.

Install Docker and Docker Compose on your machine if they are not already present
Clone the Dify repository using git clone https://github.com/langgenius/dify.git into a folder of your choice
Navigate to the cloned repository’s docker/ directory in a terminal and run docker compose up -d
Open a web browser and go to http://localhost:3000 to verify the UI loads

You'll see The Dify login screen appears in your browser, confirming the server is up and running.
Takeaway Docker Compose lets you launch multi‑container applications with a single command, simplifying deployment of complex services
Check Which directory do you run docker compose up from, and what tells you the server is up?

2.3 Ask questions that are answered from your uploaded documents

A Dify Knowledge Base that stores uploaded PDFs and provides vector‑based retrieval for chat interactions.

You will receive natural‑language answers that cite specific sections of the PDFs you added

Dify's knowledge-creation wizard after embedding, with the retrieval setting shown.

Click Create Knowledge Base and enter a name for the new collection
Select Upload Files, choose at least twenty PDF documents, and confirm the upload
Press Enable RAG to trigger chunking of the PDFs and generation of embeddings
Save the knowledge base and copy its ID for later use in a chatbot

You'll see A progress bar finishes chunking and embedding, and the knowledge base lists all uploaded files ready for retrieval
Takeaway RAG ties vector search to LLM output so responses are grounded in source material
Check What does Enable RAG start doing to the uploaded PDFs, and what tells you it has finished?

2.4 Enable citations for a chatbot built on a knowledge base

A feature that shows numbered source references for AI answers drawn from uploaded documents.

Your chatbot will display citation numbers linking back to the original PDF chunks after each response.

Open an existing chatbot app or create a new one in Dify.
In the builder, open the Features panel (available for Chatbots, Agents, and Text Generators).
Toggle on the “Citations and Attributions” option.
Ensure the app is connected to a Knowledge Base that contains uploaded PDFs.
Save the app and enter Preview mode.
Ask a question that requires information from the uploaded documents.

You'll see The AI response includes numbered citations (e.g., [1], [2]) that you can click to view the source document and chunk.
Takeaway Citations turn LLM answers into traceable, verifiable outputs by linking back to the underlying knowledge base.

2.5 Create and share a chatbot that answers with citations

A Dify Chatflow linked to the knowledge base, exposed via a shareable URL for anyone to use.

You will have a live web link where users can ask questions and receive cited answers drawn from your PDFs

A Dify chatbot switched to In Service, with its public chat URL.

Create a new Chatflow, select your knowledge base, and set a system prompt describing the bot’s tone
Add a RAG Retrieval node to the flow so each turn queries the knowledge base
Click Publish, copy the generated public URL, and open it in an incognito browser window
Enter a question that refers to one of the uploaded documents and observe the answer with its citation link

You'll see The chatbot returns a concise answer followed by a reference such as ‘(Document X, page Y)’ confirming the source
Takeaway Publishing a Dify Chatflow creates an instantly accessible AI assistant that leverages RAG to provide trustworthy, source‑backed information
Check Which node makes each turn query the knowledge base, and what does the published bot put after its answer?

2.6 Add a conversation opener with suggested questions

An opening message that greets users and offers clickable suggestions to start the chat.

Your chatbot will greet users with a custom welcome text and display suggested prompts they can click.

Open your chatbot app in Dify’s builder.
In the Features panel, locate “Conversation Opener”.
Enter a greeting message (e.g., “Welcome! How can I help you today?”).
Add up to three suggested questions using the {{variable_name}} syntax if desired.
Save the changes and publish or preview the app.
Open the chatbot UI as an end user.

You'll see The chat window shows the welcome message with clickable suggestion chips that populate the input field when clicked.
Takeaway Conversation openers improve user onboarding by guiding interactions from the first moment.

2.7 Allow users to upload files during a chat session

A file‑upload feature that lets end users attach documents, images, or audio while chatting.

Your chatbot will accept user uploads and make the file content available to the LLM for processing.

Open your chatbot app in Dify’s builder.
In the Features panel, enable “File Upload”.
Configure accepted file types (e.g., Documents) and set the maximum size limit if needed.
Save the app and go to Preview mode.
Click the upload button in the chat UI and select a PDF or image file.
Send a message referencing the uploaded file (e.g., “Summarize this document”).

You'll see The uploaded file appears as an attachment in the conversation, and the AI response includes information extracted from the file.
Takeaway File upload expands interaction modalities, allowing LLMs to reason over user‑provided content on the fly.

3You’ll know it worked 11 checkable outcomes in this chapter

✓A new member asks a question and receives an answer with citations from the PDF
✓Support tickets for routine questions are answered within seconds without human intervention
✓Researchers ask a question and receive an answer citing the correct paper title and page number
✓Writer receives a JSON brief with angle, audience, key points, and headline options after entering a keyword
✓Employees receive cited policy sections in response to a Slack query
✓Employees receive the correct leave-request link after asking a policy question
✓Agents receive answers within the ticket sidebar without leaving Zendesk
✓Tier-1 support tickets drop while users receive instant answers

11 outcomes in all — one per recipe below.

4FAQ, Tips & How-to 38

one problem, one solution, one action

Knowledge & docs3

How-to

Dify Scientist

New lab members can’t locate SOPs quickly

New lab members find protocol answers in seconds instead of hunting through shared drives or asking a senior colleague.