Paper Chatbot

This project builds a local RAG chatbot over parsed research paper chunks. It:

• Generates OpenAI embeddings for paper chunks and stores them in a local ChromaDB.
• Serves a Gradio web UI to chat with the papers.

The project's code is 100% generated by prompting GPT-5.1-Codex. It was originally designed for building a chatbot for papers listed in Awesome Graph/Transformer Fraud Detection, but one can use it to build any chatbot given pre-chunked corpus.

Setup

Use Python 3.11 or 3.12 when creating your virtual environment.

1. Create a virtual environment and install dependencies.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

2. Set your OpenAI API key (and optional overrides).

export OPENAI_API_KEY="..."

Optional overrides are listed in .env.example. If you create a .env file in the project root, it will be loaded automatically.

The chunked corpus file

Ensure data/papers.json exists before building the Chroma index. This repository does not include a corpus builder script, so generate the file externally or download a corpus of ~250 fraud-detetion-related papers parsed by MinerU.

Build the Chroma index

You can either build the chroma index using the following command with any embedding model you choose. Or you can download a prebuilt Chroma DB based on OpenAI's text-embedding-3-large. Similar to papers.json, you need to put the unzipped directory /chroma under /data.

python scripts/build_chroma.py --corpus data/papers.json --chroma-dir data/chroma

Useful flags:

• --reset to rebuild the collection from scratch.
• --skip-existing to avoid re-embedding chunks already in Chroma.
• --batch-size 100 to tune embedding batches.

Run the chatbot

python app.py

Open the local Gradio URL shown in your terminal.

Demo

Here is the screenshot of the chatbot using GPT-5.2-Chat

Below is another QA example on comparing two papers.

Click to expand

user: Can you summarize the key difference between the two papers?

1. DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector
2. Conditional Diffusion Anomaly Modeling on Graphs

assistant: Based on the provided context, the key difference between the two papers lies in how they use diffusion models for graph anomaly detection—specifically, whether diffusion is used as a data augmentation and reconstruction tool (DiffGAD) or as a direct generative modeling framework for anomaly inference (CGADM).

---

1. DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector

Core idea:

• Purpose: Transfers diffusion models from generative tasks to unsupervised anomaly detection.
• Mechanism:
  • Operates in the latent space of the graph.
  • Uses two diffusion models: one for general content and one for common content.
  • The difference between the two captures discriminative content that helps identify anomalies.
  • Employs a content-preservation strategy to maintain information consistency.
• Detection principle:
  • Nodes that are harder to reconstruct (i.e., with higher reconstruction error) are considered anomalous.
• Role of diffusion:
  • Diffusion acts as a reconstruction and feature refinement process to improve discriminative representation.
• Setting:
  • Unsupervised—no labeled anomalies are required.
• Datasets:
  • Evaluated on six real-world, large-scale datasets.

---

2. Conditional Diffusion Anomaly Modeling on Graphs (CGADM)

Core idea:

• Purpose: Introduces a conditional generative diffusion model that directly models the joint distribution of anomalies on graphs.
• Mechanism:
  • Incorporates a prior-guided diffusion process, where a pre-trained conditional anomaly estimator guides both forward and reverse diffusion.
  • Uses a prior confidence-aware mechanism to adaptively control the number of reverse denoising steps for efficiency.
• Detection principle:
  • Rather than reconstructing and comparing errors, CGADM directly generates anomaly judgments through joint distribution modeling.
• Role of diffusion:
  • Diffusion is used as a model-centric generative process, not just for data augmentation or reconstruction.
• Setting:
  • Focuses on conditional modeling—leveraging prior anomaly estimators for guided diffusion.
• Datasets:
  • Evaluated on Elliptic, Tolokers, and YelpChi datasets.

---

3. Key Conceptual Difference

Aspect	DiffGAD	CGADM
Paradigm	Reconstruction-based, unsupervised detector	Generative, model-centric conditional detector
Diffusion Role	Used to reconstruct and distill discriminative latent features	Used to model the joint distribution of anomalies directly
Guidance Mechanism	Dual diffusion models (general vs. common content)	Prior-guided conditional diffusion with adaptive denoising
Detection Basis	Reconstruction error and discriminative content	Direct anomaly probability estimation from generative model
Supervision	Unsupervised	Conditional (uses a pre-trained anomaly estimator as prior)
Goal	Enhance discriminative representation for anomaly detection	Directly generate anomaly judgments through conditional modeling

---

In summary:

• DiffGAD treats diffusion as a latent-space reconstruction and feature discrimination tool for unsupervised anomaly detection.
• CGADM redefines diffusion as a conditional generative modeling framework that directly infers anomalies by learning their joint distribution with graph structure and features.

Thus, DiffGAD = reconstruction-based detector, while CGADM = generative, model-centric anomaly modeler.