HiQA

Paper: HiQA: A Hierarchical Contextual Augmentation RAG for Massive Documents QA, arXiv:2402.01767.

HiQA provides a comprehensive toolkit for document processing, enabling the segmentation of documents into sections, enrichment with metadata, and embedding for in-depth analysis. It leverages a multi-route retrieval system to identify relevant knowledge in response to specific queries. This knowledge, along with the query, is then processed by a large language model (LLM) to generate answers. Although document processing incurs some initial costs, this investment significantly improves the quality of the results.

Usage

Ensure your environment meets the following prerequisites:

• Python version 3.9
• Install dependencies from requirements.txt using the following command: pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
• Set your OpenAI API key in the environment variables.
• To start the demo, execute: PYTHONUNBUFFERED=1 nohup streamlit run app_streamlit.py --server.port 8080 --server.address 0.0.0.0 > logs/run.log 2>&1 & Note: Before running the above command, manually create a logs directory.

Creating Your Own Dataset

To build a dataset, follow these steps:

1. Utilize the tools in the build_tool directory.
2. Begin with a PDF file that is text-extractable.
3. Step 1: Convert the PDF to a well-formatted markdown file using pdf2md, leveraging the gpt-4-turbo-preview (0125) model.
4. Step 2: Convert the markdown file into a CSV file with md2csv, organizing content into sections with hierarchical metadata, and labeling tables.
5. Step 3: Use section2embedding to append embedding vectors to sections.
6. Step 4: Place all processed CSV files into a dataset directory. Load this dataset in knowledge_client.py for querying in the app_streamlit.py demo. Note: File names and titles are processed through Named Entity Detection models to generate critical keywords, which are stored in utils.filter.critic_keywords.

Extracting and Searching Images from a PDF

For image processing:

1. In image_service, execute load, build, and commit operations to create an /indexes directory for Whoosh.
2. Use VLM (such as Ollama -> Llama:34b ) to generate descriptions for each extracted image.
3. Image searches can be conducted using app_streamlit.search_images_from_response.