🚀 PaperPilot AI

An autonomous research assistant that discovers academic papers, ranks them, indexes them, and lets you chat with them — grounded in the actual paper text, not vibes.

PaperPilot AI searches arXiv and Semantic Scholar, ranks results with a weighted scoring model (semantic similarity + citation impact + recency), downloads and indexes selected papers with LlamaIndex, and answers questions about them through a chat interface.

For the full architecture, current technical debt, and engineering conventions, see CLAUDE.md — it's the source of truth for how this repository actually works today, kept in sync with the code rather than aspirational.

---

⚡ 60-Second Product Tour

See how PaperPilot transforms research from hours into minutes.

🔍 Search Papers   →   🧠 Rank Intelligently   →   📄 Index Automatically   →   💬 Chat with Confidence

Screen.Recording.2026-07-23.054940.mp4

---

Architecture at a glance

frontend/          React 19 + Vite + TypeScript + Tailwind — search UI, workspace library, paper chat
src/app/           FastAPI backend (api.py, utils.py) — the HTTP layer the frontend talks to
src/paperpilot/    Installable Python package — search, ranking, document download,
                    LlamaIndex-backed paper chat, SQLite workspace persistence

There is also a LangGraph multi-agent pipeline (src/paperpilot/agent/, src/paperpilot/graph/) implementing a Planner → Search → Tutor → Critic self-correction loop. It's fully built and tested but not currently wired into the API — see CLAUDE.md §4 for why and what's planned.

Search & Ranking

Queries arXiv and Semantic Scholar in parallel, merges duplicate results (DOI → arXiv ID → Jaccard title similarity), and ranks the merged set with:

• Semantic similarity — cosine similarity between the query and each paper's title+abstract embedding
• Citation impact — log-normalized citation count
• Recency — exponential age decay

Document Intelligence (RAG)

Selected papers are downloaded, parsed with PyMuPDF, chunked, embedded (BAAI/bge-small-en-v1.5), and indexed with LlamaIndex (VectorStoreIndex), persisted per-paper under storage/papers/paper_<id>/ with a content-hash fingerprint so re-processing an unchanged PDF just loads the cached index. Chat is grounded via a condense_plus_context chat engine; multi-paper workspaces fan a query out across every paper's index and merge results by score.

---

Quick Start

1. Clone & set up a virtual environment

git clone https://github.com/Rg9906/Research_Assistant.git
cd Research_Assistant

python -m venv .venv
.venv\Scripts\activate     # Windows
source .venv/bin/activate  # macOS/Linux

2. Install dependencies

pip install -e ".[dev]"

This installs PyTorch, sentence-transformers, FAISS, LlamaIndex, and LangGraph — it will take a few minutes.

3. Configure environment

copy .env.example .env    # Windows
cp .env.example .env      # macOS/Linux

At minimum, set OPENAI_API_KEY in .env to enable chat/summarization. SEMANTIC_SCHOLAR_API_KEY is optional (raises the free-tier rate limit).

4. Run the test suite

pytest tests/ -v

61 tests run fully offline (all external calls mocked/stubbed). A further 9 tests in tests/test_embedder.py download the all-MiniLM-L6-v2 model from Hugging Face on first run and require network access — they're skipped automatically in network-restricted environments.

5. Run the backend

cd src
uvicorn app.api:app --reload --port 8000

src/app isn't an installed package (no __init__.py), so it must be run with src/ as the working directory / on sys.path.

6. Run the frontend

cd frontend
npm install
npm run dev

---

Code Layout

src/paperpilot/
├── core/                 # Pydantic data models: PaperMetadata, TextChunk, RetrievalResult, ...
├── config.py             # Pydantic Settings — .env-backed configuration
├── document/             # PDF downloader (validation, retries, scheme/size safety checks)
├── retrieval/             # Sentence-transformers embedder (used by the ranker)
├── search/               # arXiv + Semantic Scholar providers, dedup/merge, weighted ranker
├── services/paper_chat/  # PaperSession / PaperSessionManager — the live LlamaIndex RAG stack
├── workspace/            # SQLite-backed WorkspaceManager (workspaces, papers, mappings)
├── pipeline.py           # Facade connecting workspace + paper_chat for the LangGraph path
├── agent/                # Planner / Tutor / Critic LangChain agents (LangGraph path)
└── graph/                # LangGraph StateGraph wiring the agents together

src/app/                  # FastAPI backend consumed by the frontend
frontend/                 # React + Vite + TypeScript UI
tests/                    # pytest suite (offline-first, pythonpath=src)

---

Roadmap

See CLAUDE.md §11 for the actively-maintained roadmap. In short: stabilize the current RAG path, decide how the Planner/Tutor/Critic grounding loop reaches production traffic, then build out comparison/summarization/roadmap-generation features on a solid foundation.

---

License

MIT.