IndexGPT

中文说明

IndexGPT is a local AI research assistant workflow built around RAG + SFT + LoRA. It supports paper ingestion, retrieval-based QA, SFT data generation, LoRA training, and web chat in one project.

0. Project Identity

• Repository: https://github.com/requestsession/IndexGPT
• Maintainer: requestsession

1. Environment

• Python: 3.10+
• Node.js: 18+
• OS: Windows/Linux (current repo is Windows-friendly)
• GPU: recommended for inference/training

Install dependencies:

pip install -r requirements.txt
cd web/frontend
npm install

2. Model Placement (Important)

This project uses local model folders. A model is not a single file, but a directory containing files such as:

• config.json
• tokenizer files (tokenizer.json, tokenizer_config.json, ...)
• weights (*.safetensors / pytorch_model*.bin)

Recommended placement

Put model folders under:

workspace/rag_index/models/

Example:

workspace/rag_index/models/
  Qwen2.5-7B-Instruct/
  bge-base-en-v1.5/

Then open Settings page and select:

• Base LLM
• Embedding Model

The backend scans workspace/rag_index/models/ and lists available folders.

3. Start the App

From project root:

python web/backend/main.py

In another terminal:

cd web/frontend
npm run dev -- --host

Default ports:

• Backend: 8176 (from config.py / settings)
• Frontend: Vite default 5173

4. UI Preview

AI Assistant

Documents

Training Lab

System Settings

5. End-to-End Workflow

Step A: Upload papers

• Open Documents tab
• Upload PDF files
• Files are stored in workspace/papers/

Step B: Build RAG index

• Open Training Lab
• Click RAG Indexing -> Execute

Generated artifacts:

• workspace/rag_index/parse/*.json
• workspace/rag_index/faiss.index
• workspace/rag_index/texts.json
• workspace/rag_index/meta.json

Step C: Generate SFT data

• Click SFT Generation -> Execute
• Output file: workspace/data/sft.jsonl

Step D: Train LoRA

• Click LoRA Training -> Execute
• Output adapter path: workspace/outputs/lora/

Step E: Chat / QA

• Open AI Assistant tab
• Ask questions against indexed papers
• Optional modes:
  • Compare mode (cross-paper comparison)
  • Web mode (web snippets)

6. Equivalent CLI Commands

Run from project root:

python -u scripts/a_rag.py
python -u scripts/b_generate_sft.py
python -u scripts/c_train_lora.py

Inference module entry:

python scripts/d_infer_lora.py

7. Key Paths

• workspace/papers/: uploaded PDFs
• workspace/rag_index/: retrieval index and metadata
• workspace/data/sft.jsonl: SFT dataset
• workspace/outputs/lora/: trained LoRA output
• workspace/logs/: runtime logs
• workspace/chats/: persisted chat sessions

8. Runtime Notes

• Heavy tasks are mutually exclusive (indexing/SFT/training cannot run together)
• Chat is blocked while a heavy task is running
• Sliding-window chat history is controlled by settings (CHAT_HISTORY_ROUNDS)
• If runtime directories under workspace/ are deleted, they are auto-created again when config.py is loaded

9. Troubleshooting

No model appears in Settings dropdown

Check:

1. Model directories exist under workspace/rag_index/models/
2. Each model folder is complete (not only one weight shard)
3. Refresh Settings page / restart backend

Chat returns no sources

Usually means retrieval confidence is low. Re-check:

1. PDFs are readable text (not scanned images only)
2. Index has been rebuilt after adding PDFs
3. Query wording is specific enough

Training fails quickly

Check:

1. Base LLM path is valid
2. workspace/data/sft.jsonl exists
3. GPU/VRAM availability and dependency versions

10. Citation

If this project helps your work, please cite:

@software{indexgpt,
  title = {IndexGPT},
  author = {requestsession},
  year = {2026},
  url = {https://github.com/requestsession/IndexGPT}
}

11. License

MIT