ReadX

ReadX is an AI assistant that helps researchers read, analyze, and understand research papers. Unlike generic “chat with PDF” tools, ReadX adds:

• Structured metadata via GROBID → reliable title, author, and section extraction.
• Author context → retrieve author profiles, affiliations, and past work.
• Visual explanations → break down methods/results into diagrams and charts.
• Cross-paper analysis → compare multiple papers side by side.
• Persistent knowledge base → build your own library of papers you can query anytime.

---

Features (MVP → Future)

• ✅ Upload & parse PDFs (fallback: PyMuPDF, preferred: GROBID)
• ✅ Extract metadata (title, authors, abstract, venue, year)
• ✅ Store structured content in Postgres (papers, authors, chunks)
• ✅ Embed chunks into VectorDB for retrieval
• 🚧 LLM integration (Gemini via LangChain/LangGraph) for Q&A and synthesis
• 🚧 Visualization endpoints (e.g., plots from results section)
• 🚧 ArXiv API ingestion (auto fetch papers)
• 🚧 Multi-agent workflows (summarizer, author analyzer, visualizer)
• 🚧 Slack/Discord bot integration

---

Architecture

See docs/architecture.md for system diagrams. Key components:

1. Ingestion

   • GROBID XML parsing (structured sections, references)
   • PyMuPDF heuristics (fallback when GROBID is unavailable)

2. Storage

   • Metadata & relationships → Postgres (papers, authors, chunks)
   • Embeddings → ChromaDB / Weaviate

3. Orchestration

   • LangGraph workflows for question answering, author analysis, visualization
   • Gemini as the preferred LLM backend

4. API/UI

   • FastAPI endpoints (/papers, /query, /author, /visualize)
   • Streamlit chat + visualization frontend

---

Database Schema

┌─────────────────────┐       ┌──────────────────────┐
│       papers        │       │       authors        │
├─────────────────────┤       ├──────────────────────┤
│ id (PK)             │       │ id (PK)              │
│ filename            │       │ name (UNIQUE)        │
│ title               │       └──────────────────────┘
│ abstract            │               ▲
│ year                │               │
│ venue               │               │
│ path                │               │
│ created_at          │               │
└─────────┬───────────┘               │
          │                           │
          ▼                           │
┌───────────────────────┐   ┌────────────────────────┐
│    paper_authors      │   │     paper_chunks       │
├───────────────────────┤   ├────────────────────────┤
│ paper_id (FK→papers)  │   │ id (PK)                │
│ author_id (FK→authors)│   │ paper_id (FK→papers)   │
└───────────────────────┘   │ section                │
                            │ chunk_index            │
                            │ content                │
                            └────────────────────────┘

• papers → stores global metadata.
• authors → unique author names (linked across papers).
• paper_authors → many-to-many relation between papers & authors.
• paper_chunks → sectioned + chunked content for embedding.

---

Tech Stack

• LLM Orchestration → LangChain, LangGraph
• LLM Provider → Gemini (2.5 Flash / Pro, configurable)
• Backend → FastAPI
• Vector Store → ChromaDB / Weaviate
• Database → Postgres (persistent metadata & author graph)
• UI → Streamlit + Plotly (charts & visualizations)
• Infra → Docker, GitHub Actions, Kubernetes (future)

---

Setup

# clone repo
git clone https://github.com/<your-username>/ReadX.git
cd ReadX

# setup virtual environment
python3 -m venv .venv
source .venv/bin/activate

# install dependencies
pip install -r requirements.txt

# (optional) start GROBID (Docker)
docker run -it -p 8070:8070 lfoppiano/grobid:0.7.2

# start postgres (if local)
brew services start postgresql@15

# run backend
uvicorn app.main:app --reload

# run UI
streamlit run ui/streamlit_app.py

---

Next Steps

• Add TEI body chunking (structured section-level chunks from GROBID)
• Improve author disambiguation (affiliations, emails)
• Add LLM-powered synthesis (Gemini) into /query/ask
• Expand visualization layer (method diagrams, results plots)

---