A WebUI that makes ChatGLM read your PDFs before answering

What it does Takes the standard retrieval-augmented generation stack—LangChain, Chinese-leaning LLMs like ChatGLM-6B and BELLE-LLaMA, and embedding models such as text2vec-large-chinese—and wraps it in a WebUI. You upload txt, docx, md, or pdf files, and the model grounds its answers in that local knowledge rather than relying on training-data hallucinations. The authors also host live demos on HuggingFace, ModelScope, OpenXLab, and Baidu’s AIStudio, so you can try before you buy the GPU time.

The interesting bit The project is refreshingly honest about being early-stage and built atop the upstream langchain-ChatGLM framework. Rather than pretending to reinvent the wheel, it focuses on accessibility: bundling support for domestic embeddings like ERNIE and SimBERT, and chasing compute donations from Chinese platforms such as OpenI and ModelScope instead of defaulting to Western cloud hosts.

Key highlights

• Supports a wide model menu: ChatGLM-6B (including int4/int8 quantized variants), Vicuna, BELLE-LLaMA, and InternLM.
• Embedding options include text2vec-large-chinese, ernie-3.0-base-zh, and simbert-base-chinese, among others.
• Accepts txt, docx, md, and pdf as knowledge sources.
• Runs fully offline if you have the hardware, but also offers hosted playgrounds on HuggingFace, ModelScope, OpenXLab, and AIStudio.
• Built on Python ≥3.8.1 and explicitly credits its upstream inspirations.

Caveats

• The README states the project is in its early stage with plenty of optimization still to do.
• Most of the heavy lifting is done by LangChain, ChatGLM-6B, and the upstream langchain-ChatGLM project; this repository is primarily the WebUI and integration layer.

Verdict Worth a look if you need a quick, Chinese-LLM-friendly RAG interface for local documents and prefer clicking to coding. Pass if you’re already running a custom LangChain pipeline and don’t need another abstraction layer.