Research Explorer

A dark-themed research web app built for a principal data scientist — streams the latest GenAI and ML papers live from arXiv across 8 curated topic feeds. Browse, search, annotate, and save notes on papers that matter to your work.

---

What This App Does

Research Explorer solves the problem of staying current with fast-moving GenAI research without spending hours on Google Scholar or Twitter. It gives you a single, always-up-to-date view of the papers most relevant to applied ML work — especially in production environments like financial services.

The core workflow:

1. Browse — pick a topic from the sidebar (e.g. Multi-Agent Systems) and instantly see the 30 most recent arXiv papers, sorted by submission date
2. Search — run a free-text search across arXiv titles and abstracts when you need something specific
3. Read — click Read on any paper card to open the split-view reader: PDF on the left, your notes on the right
4. Annotate — write notes alongside the paper; they auto-save as you type and are stored as Markdown files on your local machine
5. Review — revisit all your annotated papers in the Saved Notes view, with previews and metadata intact

---

Key Features

Paper Discovery

• 8 curated topic feeds — each with a hand-tuned arXiv query covering the most important terms for that topic
• Live arXiv API — every fetch hits arXiv directly; papers are sorted by submission date with no stale intermediate index
• Date window filter — scope results to the last 30 days, 90 days, 6 months, or 1 year
• Full-text search — searches arXiv title and abstract simultaneously across all CS and quantitative finance categories
• Refresh on demand — a Refresh button bypasses the 2-hour cache to pull the absolute latest papers at any time

Reading Experience

• Expandable abstracts — read the full abstract inline on the card without navigating away
• NEW badge — papers submitted in the last 7 days are highlighted so you can spot the freshest work at a glance
• Split-view PDF reader — click Read to open a dedicated page with the PDF on the left and your notes on the right; drag the divider to resize the panels to any ratio
• Native PDF tools — the PDF renders in the browser's built-in viewer, so text selection, copy, and browser highlight tools all work exactly as normal
• Copy arXiv ID — one-click copy button for citing or sharing a specific paper

Paper Annotation

• Split-view note editor — notes panel sits alongside the PDF so you can write while you read, without switching windows
• Inline note editor — click 📝 Add Note on any paper card for a quick note without leaving the main feed
• Auto-save — notes in the split-view save automatically 2 seconds after you stop typing; a status indicator shows Unsaved → Saving… → Saved ✓
• Keyboard shortcut — press ⌘↵ (Mac) or Ctrl↵ (Windows/Linux) to save immediately in either editor
• Markdown persistence — each note is saved to notes/{paper-title}.md on your local machine, containing the full paper metadata (arXiv ID, authors, categories, links) plus your notes
• Edit and delete — reopen any note to update it; the file is overwritten in place
• 📝 NOTE badge — cards with existing notes show an amber badge so you can see what you've already reviewed
• Saved Notes view — a dedicated sidebar section lists all annotated papers with note previews, save timestamps, and filenames

Saved Notes Format

Each note is saved as a readable, portable Markdown file:

# Attention Is All You Need

**arXiv ID:** `1706.03762`
**Authors:** Ashish Vaswani, Noam Shazeer, ...
**Published:** 2017-06-12
**Categories:** cs.CL, cs.LG
**Abstract:** https://arxiv.org/abs/1706.03762
**PDF:** https://arxiv.org/pdf/1706.03762

---

## Notes

Foundational transformer paper. Self-attention mechanism eliminates
recurrence entirely — key insight for our document classification work.
Revisit the multi-head attention section before the architecture review.

---

*Last saved: 2026-04-29 14:30*

Files are stored locally in notes/ and are compatible with Obsidian, VS Code, Typora, or any Markdown viewer.

---

Topics

Icon	Topic	What it covers
⚙️	LLM Fine-Tuning	LoRA, QLoRA, PEFT, DPO, RLHF, instruction tuning, SFT
📊	Agentic Evaluation	Benchmarks and evaluation frameworks for AI agents
🤝	Multi-Agent Systems	Orchestration frameworks, agent coordination and communication
🔍	RAG & Retrieval	Retrieval-augmented generation, GraphRAG, knowledge retrieval
🧠	Reasoning & Planning	Chain-of-thought, ReAct, tree-of-thought, structured inference
🛡️	AI Safety & Alignment	Safety, alignment, red-teaming, adversarial robustness
💹	AI in Finance	LLMs for risk management, fraud detection, credit scoring, trading
🏗️	Foundation Models	LLM architecture, scaling laws, pre-training, model compression

---

Prerequisites

• Python 3.11+
• uv (curl -LsSf https://astral.sh/uv/install.sh | sh)

---

Setup

# 1. Clone / enter the project
cd research-app

# 2. Install dependencies (creates .venv automatically)
uv sync

---

Running

uv run uvicorn main:app --host 0.0.0.0 --port 8765

Open http://localhost:8765 in your browser.

For development with auto-reload:

uv run uvicorn main:app --host 0.0.0.0 --port 8765 --reload

---

API Endpoints

Method	Path	Description
GET	/	Main UI
GET	/pdf/{arxiv_id}	Split-view PDF reader + notes for a paper
GET	/api/topics	All topic configs and cache status
GET	/api/papers/{topic_id}	Papers for a topic (?days=90&refresh=false)
GET	/api/search	Full-text arXiv search (?q=...&days=90)
POST	/api/notes	Save a note for a paper
GET	/api/notes	List all saved notes with previews
GET	/api/notes/{arxiv_id}	Get note content for a specific paper
DELETE	/api/notes/{arxiv_id}	Delete a note

/api/papers/{topic_id} parameters

Param	Default	Description
days	90	How many days back to search
refresh	false	true to bypass the 2-hour cache

Valid topic_id values

llm_finetuning, agentic_eval, multi_agent, rag, reasoning, ai_safety, ai_finance, foundation_models

---

Configuration

Key constants in main.py:

Constant	Default	Description
CACHE_TTL	7200	Paper cache TTL in seconds (2 hours)
REQUEST_DELAY	4.0	Minimum gap between arXiv API calls (rate limit)

arXiv enforces ~1 request per 3 seconds. The 4-second delay gives safe headroom.

---

Project Structure

research-app/
├── main.py                  # FastAPI app, arXiv client, notes API, caching
├── pyproject.toml           # Project metadata and dependencies
├── .pre-commit-config.yaml  # ruff lint + format hooks
├── templates/
│   ├── index.html           # Main UI — topic feeds, search, paper cards
│   └── pdf_view.html        # Split-view PDF reader + notes editor
├── notes/                   # Saved note files — gitignored, local only
│   ├── .index.json          # arXiv ID → filename lookup map
│   └── *.md                 # One Markdown file per annotated paper
└── README.md

---

Tech Stack

Layer	Technology
Backend	FastAPI + uvicorn
HTTP client	httpx (async)
Templates	Jinja2
Frontend	Alpine.js + Tailwind CSS (CDN, no build step)
Paper source	arXiv API (free, no auth required)
Code quality	ruff via pre-commit
Package manager	uv

---

Keyboard Shortcuts

Key	Action
/	Focus the search bar from anywhere on the page
Enter	Submit search
⌘↵ / Ctrl↵	Save note (inline editor or split-view reader)

---

Notes

• Papers are cached in memory per topic + date window. Cache clears on server restart.
• The arXiv API is free and requires no API key or account.
• The notes/ directory is gitignored — your annotations stay local and private.
• Notes files are plain Markdown and open in any editor (Obsidian, VS Code, Typora, etc.).