REQUEST GUIDANCE IN CREATING AN AI AGENT TO RECOMMEND UNDERGRADUATE COURSES USING RAGFLOW/OLLAMA

[staann]

I am developing a website that has a virtual assistant to recommend subjects from the University I study at. This project is part of the "Software Development Methods" course at the Universidade de Brasília.

I webscraped the subjects from all departments at my University to provide ragflow with the necessary data.

The data for each subject is organized as follows:

• Subject name:
• Subject code:
• Syllabus:

I uploaded the LLM "OLLAMA" to be the LLM for my AI Agent. However, I am having difficulties. Because my AI Agent, when I send it study content so it can search the database for classes that make sense to take, it returns some classes that partially exist (for example, the name is almost correct, because it always adds an extra name that does not exist to the class name) and sometimes it returns classes that, despite having a similar name in the subject name, have nothing to do with it, for example (input: Neural Networks, output: Local Networks - class 123, instead of returning a subject that has to do with Artificial Intelligence).

I find the process of creating prompts and optimizing the flow of my AI agent very confusing, and I can't learn much.

I would appreciate any help that anyone can provide me, I would be very grateful.
If you want to help me, please contact me

[dosubot[bot]]

Hi @staann! I'm Dosu and I’m helping the ragflow team.

To improve your AI agent for recommending undergraduate courses using RAGFlow and OLLAMA, here are some targeted steps and troubleshooting tips based on your setup and the issues you described:

1. RAGFlow and OLLAMA Integration
Make sure you are running both RAGFlow and OLLAMA in your local environment (not the demo site), so they can communicate over the network. If you use Docker, set the OLLAMA_HOST environment variable to the private IP address of the server running Ollama, and ensure the correct ports are exposed. For Windows, use http://host.docker.internal:11434 as the connection URL; for Linux, you may need to add extra_hosts in your Docker Compose file. If the Ollama chat model does not appear in RAGFlow, try entering any random string in the API Token field as a workaround for a known bug. Always ensure the model name and type in RAGFlow match those you pulled with Ollama reference reference.

2. Data and Tagging
Accurate recommendations depend on well-structured data and tagging. Make sure your knowledge base has at least one properly configured tag set, and specify it in the configuration. Prepare tag tables with clear descriptions and tags, avoid overlapping tags, and regularly update them to reflect changes in your dataset or user queries. This helps the retrieval system match user questions to the most relevant courses reference.

3. Retrieval and Similarity Settings
The Retrieval component uses a similarity threshold (default 0.2) and a keyword similarity weight (default 0.5) to determine which courses are relevant. If you are getting partially correct or irrelevant results, try adjusting these parameters. Lowering the similarity threshold can make retrieval stricter, while increasing it can make it more permissive. Also, ensure your embedding models are consistent across all knowledge bases reference.

4. Prompt Engineering and Categorization
Prompts are critical for guiding the LLM to produce accurate recommendations. The Categorize component uses prompts with category descriptions and examples to classify user questions. Make sure your prompts are clear, include relevant examples, and describe each category (e.g., course area or department) in detail. The Generate component also allows you to set parameters like temperature, max tokens, and penalties, which can affect output quality. Experiment with these settings to see what yields the most accurate results reference reference.

5. Troubleshooting Incorrect Recommendations
If the agent returns classes that do not exist or are only partially correct, check for:

• Incomplete or ambiguous data chunks (make sure each subject entry is accurate and complete).
• Poorly configured or missing tags.
• Retrieval settings that are too permissive.
• Prompts that are too vague or do not provide enough context/examples.

There is no dedicated "course recommendation" module in RAGFlow; recommendations are handled by configuring the pipeline, knowledge base, and prompts. The system is highly data- and config-driven, so careful setup is essential reference.

If you encounter specific errors (like KeyError: 'llm'), you may need to patch the code as described in the issues reference.

If you need more concrete examples or run into a specific error message, let me know the details and I can help you troubleshoot further.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[ZhenhangTung]

@staann welcome to join our discord community and we can have some discussions there.