Tswoen/Paper-Agent

Paper‑Agent – An AI‑powered research assistant

What it does

• You type a research topic and the system automatically searches for papers, reads them, analyses the findings, and drafts a full‑length literature review you can follow step‑by‑step.
• The whole pipeline is built from a set of specialized agents (search, read, analyse, outline, write) that talk to large‑language‑model (LLM) providers such as OpenAI or Anthropic.
• Results, progress and intermediate artefacts are shown in a web‑based workbench that updates in real time via Server‑Sent Events and stores everything in a local SQLite + file system so you can pause and resume later.

Key technical pieces

Layer	Technology	Role
Runtime	Python 3.12+, uv package manager	Runs the backend services
API	FastAPI (Uvicorn)	Exposes REST endpoints and SSE streams
Workflow engine	LangGraph (typed‑dict shared state)	Orchestrates the multi‑agent pipeline
Agents	Custom Python agents (Search, Read, Analyse, WritingOutline, Writing)	Each handles a distinct stage of the research workflow
LLM integration	OpenAI‑compatible and Anthropic‑compatible providers	Generates prompts, extracts information, writes text
Paper sources	arXiv, OpenAlex, Semantic Scholar connectors	Retrieves metadata and PDFs
PDF processing	pypdf → Markdown → text chunking	Turns PDFs into searchable text
Vector store	ChromaDB	Stores chunk embeddings for fast similarity search
Persistence	SQLite + JSON/Markdown files	Saves sessions, intermediate results, and final drafts
Frontend	Vue 3 + TypeScript + Vite	Interactive UI showing progress, model configuration, and chat‑style session view
Package management	uv (Python) & npm (frontend)	One‑command install of all dependencies

How you use it

1. Install – run uv sync (Python deps) and npm run front:install (frontend deps).
2. Configure – create config/model.json (or use the web UI) with your LLM provider URL, API key, and model “tiers” for each agent.
3. Start – uv run python main.py launches the FastAPI server (http://127.0.0.1:8000/docs for API docs).
4. Run the UI – npm run front:dev (or front:dev:network for LAN access) opens the Vue app at http://127.0.0.1:5173/.
5. Create a session – enter a research topic, watch the SSE‑driven progress bar as the system searches three sources, reads abstracts, downloads PDFs, chunks them, analyses sub‑topics, builds an outline, and writes each section.
6. Save & resume – all state is persisted; you can close the browser and later continue where you left off.

Why it matters

• End‑to‑end automation: replaces the manual loop of searching, reading, note‑taking, outlining, and drafting.
• Cost awareness: each stage can use a different model tier, and the UI shows token usage per step.
• Transparency: real‑time SSE updates let you see exactly what the system is doing and intervene if needed.
• Extensible: built on LangGraph, so new agents or custom prompts can be added without rewriting the whole stack.

Getting involved

• Issues and pull requests are welcomed. Tests live in test/ and can be run with uv run python -m unittest discover -s test -v.
• Front‑end changes should pass npm run front:build before merging.
• The project is open‑source under the repository https://github.com/Tswoen/Paper-Agent.

---

Paper‑Agent aims to make literature review faster and more systematic by letting LLMs handle the heavy lifting while keeping the whole process visible and controllable.