RAGFlow 教學:開源 RAG 引擎怎麼裝、怎麼餵文件、怎麼讓 AI 不亂講

Introduction: When AI Misreads PDFs

I've come across a real case where a company wanted to build an internal knowledge base by feeding over 300 product manuals in PDF format into an RAG system. However, the AI would incorrectly connect numbers in tables to other text segments and mistake copyright notices at the end of pages for actual content references. When asked about the power of a specific model, it would confidently provide an incorrect answer that actually belonged to a different model. The engineer in charge sighed, saying the problem wasn't with the model being dumb, but rather that the files weren't being properly read into the system.

This is where RAG truly faces challenges - not in connecting models, but in accurately segmenting complex files with tables and scanned documents into pieces that the model can handle. RAGFlow specializes in this aspect. This article will guide you through setting it up, feeding it files, and enabling it to provide sources for its answers.

What is RAGFlow

RAGFlow is an open-source RAG (Retrieve, Augment, Generate) engine developed by InfiniFlow, with its code publicly available on GitHub. If you're still unsure what RAG is, it's best to start with our What is RAG article - simply put, it's a technology that allows AI to search your data repository before answering questions, rather than relying on memory or making things up.

What sets RAGFlow apart from other RAG tools is its "DeepDoc" deep document understanding capability. Unlike general tools that simply extract text from PDFs, which can become disorganized when the layout is complex, RAGFlow understands the structure of the document - identifying titles, tables, paragraphs, and even handling scanned documents (image-based PDFs) through OCR. It supports a wide range of formats including Word, presentations, Excel, plain text, images, scanned documents, structured data, and web pages.

Another key feature is traceable references. When answering, it will mark the source of the information, allowing you to see which part of which document the answer comes from, significantly reducing the chance of being misled by the AI. It also integrates with platforms like Dify that are working on agent orchestration, supporting the creation of visual workflows that combine RAG, tools, and MCP.

Use Cases

• Enterprise Internal Knowledge Base: Feed product manuals, SOPs, and contract templates into it, and employees can ask questions in natural language, with answers provided along with their sources.
• Customer Service FAQ Backend: Connect it with frequently asked questions and technical documents to serve as the brain for chatbots.
• Research and Analysis: For complex documents like legal cases, financial reports, and academic papers that require precise references, DeepDoc's strengths are particularly useful.
• Personal Second Brain: Throw in your accumulated notes and e-books, and it becomes a conversational knowledge base.

To understand how to set up an RAG system as a whole, you can refer to our RAG Implementation Guide alongside this article.

Getting Started: First-Time Setup

RAGFlow is deployed using Docker, and while the process isn't difficult, it does have some system requirements. Let's take a look.

1. Confirm System Requirements

The official recommendation is for a CPU with at least 4 cores, at least 16GB of RAM, and at least 50GB of hard drive space, with Docker 24 or above and Docker Compose v2.26 or above. The RAM requirement is crucial - it needs to run a vector search engine, and insufficient RAM will cause it to freeze, so don't try to run it on a small machine with only 4GB of RAM.

2. Pull the Project and Start

After cloning the project from GitHub, navigate into the docker directory and start it:

bash
git clone https://github.com/infiniflow/ragflow.git
cd ragflow/docker
docker compose up -d

The first run will download several images, so grab a cup of coffee and wait. Once it's done, use docker ps to confirm that all containers are up and running.

3. Open the Backend and Set Up the Model

After the service starts, open http://your-host-IP (default port 80) in your browser. Register an account and log in, then go to the settings to fill in the API key for your model - this step cannot be skipped. RAGFlow does not come with a model; you need to tell it which LLM (for answering) and which embedding model (for converting text into vectors) to use. You can fill in OpenAI, cloud models from various providers, or even connect to a local Ollama.

4. Build a Knowledge Base and Feed Files

Create a Knowledge Base and upload your PDFs and Word documents into it. The key part comes next: after uploading, you need to select a chunking template (chunk method). RAGFlow provides multiple templates corresponding to different file types - general documents, academic papers, books, legal documents, presentations, tables, Q&A pairs, etc. Choosing the right template significantly affects the quality of the chunking. After selecting the template, press parse, and it will run DeepDoc to break down the files into segments.

5. Check the Chunking Results

This is one of RAGFlow's most useful features: after parsing, it will visualize the chunking results for you, allowing you to see what each segment looks like and where it corresponds to in the original text. If you find that the chunking is poor - for example, if tables are cut off - you can manually adjust it. Don't be afraid of the trouble; this step directly affects the accuracy of the answers later on.

6. Build a Chat Assistant and Test Q&A

Once the chunking is satisfactory, build a Chat and hang the knowledge base on it, and you can start asking questions. When it answers, it will attach the source references, and clicking on them will jump to the original text segment.

Advanced Tips

Use templates according to file types: Don't put legal contracts and product briefs into the same knowledge base using the same template. The layouts are too different, and the chunking logic isn't the same. Build separate bases and choose appropriate templates for each, and the accuracy will significantly improve.

Make good use of hybrid search: RAGFlow performs both vector search and BM25 keyword search, followed by re-ranking. Pure vector search often fails to catch "exact nouns, material numbers, article numbers," which keyword search can complement. It's enabled by default, but you can adjust the weights.

Connect to Ollama to run local models: For those concerned about data privacy and not wanting to send company files to the cloud, you can point both the LLM and the embedding model to a local Ollama, running the entire suite offline. The trade-off is that you'll need better hardware.

Use API to integrate into your system: RAGFlow provides an API that you can use as a backend search layer, with your own interface on the frontend, or integrated into existing customer service systems. For more complex workflows, you can also use its agent orchestration feature, similar in concept to AI Agent Development.

Common Errors and Notes

• Insufficient RAM: 16GB is the minimum, not a recommendation. Running below this will cause containers to crash repeatedly, a common issue for newcomers.
• Forgetting to set the embedding model: Many people only set up the LLM for answering and forget the embedding model, resulting in files not being correctly vectorized, and thus answering poorly. Both need to be set.
• Carelessly choosing or not checking chunking templates: Using the default template for a complex PDF and then complaining about poor answers - the problem likely lies in the chunking. You must look at the visualization results and adjust if they're bad.
• Believing references guarantee 100% accuracy: Traceable references allow for verification, not a guarantee of correctness. References reduce the risk of hallucinations, but models can still misread segments. For critical scenarios, please manually verify the sources. To understand more about why AI might provide incorrect information, see What is AI Hallucination.
• Not considering data privacy: Using cloud models means your file segments will be sent to the model providers. For sensitive data, please use local model solutions.

TheAI Academy Review

Among the many tools that can connect to RAG, most focus on how to connect models or store vectors, assuming files can be cleanly read into the system. In practice, file parsing is where 90% of projects fail. RAGFlow focuses its efforts on DeepDoc and chunking visualization, directly addressing this dirtiest of jobs. For teams wanting to build enterprise knowledge bases and being driven crazy by tables and scanned documents, this focus is very appropriate.

After doing RAG for a long time, you'll understand a phrase: the quality of answers isn't won by how strong the model is, but by whether the files have been properly read into the system. RAGFlow bets on this, and it's the right bet.

Note that it's an engineering-oriented tool that requires Docker knowledge, a decent machine, and isn't a cloud service that you can simply install and use. If you're an individual looking to simply try RAG, the threshold might be too high; but if you're looking to self-host, control, and handle complex files, it's worth spending an afternoon to set up and play with.

Data Sources

• RAGFlow Official Website
• RAGFlow GitHub