Research-It is a fully local Retrieval-Augmented Generation (RAG) system for research papers.
It lets you ingest PDFs, URLs, or directories of papers → build vector indexes → and chat with them via a Streamlit UI backed by Ollama LLMs.

Run everything offline on your own machine — no API keys, no cloud costs.

• 📄 Index Any Research Source
  • Direct PDF files (local or from the web)
  • HTML pages (arXiv, NeurIPS, ResearchGate, etc.)
  • Entire folders of papers (./data)
• 🧩 Chunking + Embeddings
  • Smart text chunking with overlap
  • sentence-transformers for dense embeddings
  • hnsw index backend (FAISS-compatible, efficient)
• 🔍 Fast Local Retrieval
  • LEANN (leann) powers vector search with pruning + quantization
• 🤖 Local LLM
  • Ollama for running quantized models (llama3.2:1b, 3b, etc.)
  • No GPU? Use CPU or small quantized models
• 🎨 Streamlit UI
  • Upload a paper or enter a URL → index instantly
  • Chat about your documents with sources and context shown
  • Sidebar lets you pick indexes, tweak chunk size, overlap, retrieval Top-K, context length, and model
• 🔒 Privacy by Design
  • No data leaves your machine
  • No API keys, no cloud dependencies

• Python 3.10+
• LEANN → embeddings, HNSW/CSR vector indexes
• sentence-transformers → embeddings (default: facebook/contriever)
• Ollama → local quantized LLMs
• Streamlit → simple and responsive chat UI
• BeautifulSoup + readability → web text cleaning
• PyMuPDF → PDF parsing
• Typer → CLI utilities

We provide a one-command setup script for Linux/macOS and Windows.
It installs Python deps, Ollama, pulls models, and sets up project folders.

git clone https://github.com/spice14/research-it.git
cd research-it

# Run setup
bash scripts/setup.sh

# Activate virtualenv
source .venv/bin/activate

# Launch the app
streamlit run app_streamlit.py

git clone https://github.com/spice14/research-it.git
cd research-it

# First time only
Set-ExecutionPolicy -Scope CurrentUser Bypass -Force

# Run setup
.\scripts\setup.ps1

# Activate venv
.\.venv\Scripts\Activate.ps1

# Launch the app
streamlit run app_streamlit.py

---

• From a PDF:

python -m src.index.build_any \
  --index-path ./indexes/my-paper.leann \
  --file ~/Downloads/some_paper.pdf

• From a URL:

python -m src.index.build_any \
  --index-path ./indexes/my-paper.leann \
  --url "https://arxiv.org/pdf/2307.09218v3.pdf"

• From a folder of papers:

python -m src.index.build_any \
  --index-path ./indexes/my-library.leann \
  --dir ./data

This generates:

./indexes/<name>.index
./indexes/<name>.meta.json
./indexes/<name>.passages.idx
./indexes/<name>.passages.jsonl

Launch the app:

streamlit run app_streamlit.py

Features in the sidebar:

• Pick an existing index or type a path
• Build a new index from URL or PDF upload
• Select Ollama model (llama3.2:1b recommended for small GPUs)
• Adjust retrieval params (chunk size, overlap, Top-K, num_ctx)

Ask questions in the chat box — you’ll get an answer with retrieved sources.

• Low VRAM (≤4GB GPU)?

  • Use llama3.2:1b or another quantized small model
  • Reduce num_ctx (e.g., 1024–1536)
  • Lower Top-K (3–4)

• No GPU?

  • Ollama runs CPU-only models too, just slower

• Avoid CUDA OOM: Export before running:

  export LEANN_EMBED_BATCH=4
  export PYTORCH_CUDA_ALLOC_CONF="max_split_size_mb=64,expandable_segments:True"

research-it/
├── app_streamlit.py       # Streamlit UI
├── scripts/
│   ├── setup.sh           # Linux/macOS setup
│   └── setup.ps1          # Windows setup
├── src/
│   ├── index/             # build_any, build_from_url, etc.
│   ├── ingest/            # PDF, HTML, file loaders
│   └── rag/               # CLI chat pipeline
├── indexes/               # saved vector indexes
├── data/                  # cleaned paper text dumps
└── uploads/               # optional file uploads

• Support more embedding backends (bge-m3, InstructorXL, etc.)
• Add incremental indexing / updates
• Hybrid retrieval (BM25 + dense)
• Docker image for one-shot deploy
• Offline model bundling for air-gapped use

• LEANN — blazing fast vector store
• Ollama — local quantized LLMs
• SentenceTransformers — embeddings
• Streamlit — UI
• Inspired by the need to read and truly understand papers locally ✍️