Skip to content

Latest commit

 

History

History
133 lines (82 loc) · 7.83 KB

user_guide_en.md

File metadata and controls

133 lines (82 loc) · 7.83 KB

Introduction and Usage Guide for WebUI

WebUI uses a two-section layout. The top of the page is for Global Settings, while the bottom is for Configuring Each Stage of the Process. Global Settings includes the Model Management and Knowledge Management tabs. Process configuration covers the stages of Data Construction, Training, Evaluation, and Inference.

When using WebUI, the following steps are typically followed:

  1. Select and load the appropriate model on the Model Management page;
  2. Switch to the Knowledge Management page to upload documents and complete the knowledge base construction;
  3. Go to the Chat/Inference page to choose the suitable workflow for interaction.

Below, we will introduce the specific usage methods for each section of WebUI.

img

Model Management

Model Management includes three types of models: LLM, Embedding, and Reranker. Currently, it supports both local loading and API loading. In the future, more models will be supported, and a one-click microservice launch feature will be provided to optimize the model management process.

Users can download the required models from model repositories like ModelScope and Hugging Face, deploying them flexibly to meet the demands of different application scenarios.

Inference Supported Models
Model Type Local Model Microservice
LLM Models supported by VLLM MiniCPM-V and other models supported by VLLM
Embedding bge, minicpm-embedding
Reranker bge-large-reranker

Knowledge Base Management

Knowledge Base Management is composed of File Management and Knowledge Base Management. Currently supported file types include PDF and TXT. The usage process is as follows:

  1. Upload File: Users must first upload the files and load the corresponding Embedding model;
  2. File Selection: After selecting the required files, configure the basic parameters of the knowledge base;
  3. Knowledge Base Construction: After completing the parameter configuration, you can build the knowledge base. The Model Name is the only identifier for determining whether the knowledge base can be mixed-use and must be customized by the user.

Note: Both File Management and Knowledge Base Management are maintained by their respective CSV tables, recording the relevant basic information.

The constructed knowledge base will include the following three parts for downstream algorithm use:

  • Index (used for retrieving top_k results);
  • Org_Files (original files);
  • Chunk_Files (chunked files).
Illustration:

img

Data Construction

On the Data Construction page, we provide data construction methods for various algorithms, currently mainly supporting downstream tasks such as Training and Evaluation. In the future, functionality will be further expanded, such as supporting the automatic construction of knowledge base data.

Illustration:

img

Training

The training page mainly supports the training of LLM and Embedding models, with only partial core parameter configurations available on the interface. LoRA is primarily used for LLM training, allowing users to flexibly input specific parameters through configuring YAML files.

Main features include:

  • Dataset preview (resource/dataset/train_dataset);
  • SFT/DPO training to meet various fine-tuning needs;
  • Embedding model training for efficient vector representation;
  • Support for other training algorithms (e.g., UltraRAG-KBAlign);
  • LoRA parameter merging to enhance the flexibility of model training.
Illustration:

img

Evaluation

The evaluation page allows for one-click evaluation of multiple processed datasets, covering both retrieval and generation modules. Additionally, users can directly calculate various metric scores for already generated result files, provided that the file formats comply with the specifications.

Main features include:

  • Dataset preview (resource/dataset/test_dataset) to check data integrity;
  • Retrieval metrics evaluation to measure model retrieval precision and recall performance;
  • Generation metrics evaluation to assess the quality and consistency of generated text;
  • Integrated retrieval+inference+evaluation, supporting API and vLLM based workflows;
  • Direct evaluation for already generated and compliant result files to quickly obtain evaluation results.
Illustration:

img

Chat/Inference

The chat page supports multiple workflows, including Vanilla RAG, UltraRAG-VisRAG, UltraRAG-KBAlign, and UltraRAG-Adaptive-Note, with plans to continuously expand more workflows to further enhance the system's flexibility and applicability. Before using the chat function, users need to complete the following steps:

  1. Select an appropriate model and build a knowledge base;
  2. After completing the above steps, select the corresponding workflow and knowledge base to initiate the chat process.
Basic Usage

Engaging in dialogue using WebUI and VanillaRAG requires the following steps:

  1. Model Configuration: In the Model Configuration Panel, select the default model path and sequentially load the large model, embedding model, and rerank model.
  2. Knowledge Configuration:
    1. In the Knowledge Configuration Panel, upload a file (e.g., PDF);
    2. Select the uploaded file and use the default parameters to build the knowledge base.
  3. Chat Experience: In the Experience Panel, select Chat/Inference, check the knowledge base and workflow to start interacting in the chat window.

For detailed operational instructions, please refer to the demonstration video.

VisRAG

Engaging in dialogue using WebUI and VisRAG requires the following steps:

  1. Load LLM Model: In the Model Configuration Panel, enter the MiniCPM-V model path, select the GPU, and then click Load to complete the LLM loading.
  2. Load Embedding Model: In the Model Configuration Panel, enter the VisRAG-Ret model path, select the GPU, and then click Load to complete the embedding model loading.
  3. Knowledge Base Construction:
    1. In the Knowledge Configuration Panel, upload a file (e.g., PDF) or check a previously uploaded file;
    2. Select the uploaded file and use the default parameters to build the knowledge base.
  4. Chat Experience: In the Experience Panel, select Chat/Inference, check the knowledge base and workflow to start the interaction in the window.

For detailed operational instructions, please refer to the demonstration video.

VisRAG/Adaptive-Note Evaluation

Refer to the demonstration video for evaluation operations, which need to follow these steps:

  1. Select Workflow: Select the workflow to be tested.
  2. Select Knowledge Base: Choose the knowledge base needed for testing.
  3. Configure Evaluation Parameters: Check the evaluation set and metrics; specify the output path for results.
  4. Run Command: Click Preview Command to get the corresponding run command; click Execute Command to perform the evaluation on the frontend and preview execution results in real-time in the window below.

Suggestion: Since evaluation time may be long, closing the frontend might interrupt the command execution. It is recommended to copy the command obtained and run it locally on the server to ensure stability and completeness of the evaluation process.

img