Curator Agent

An agentic content curation tool that turns any source material — URLs, files, or pasted text — into a structured learning challenge, then evaluates the learner's response with an LLM judge.

---

Prerequisites

Ollama running locally with two models pulled:

ollama pull llama3
ollama pull nomic-embed-text

Docker for Qdrant and Neo4j:

docker compose up -d

This starts:

• Qdrant at localhost:6333
• Neo4j at localhost:7474 (HTTP) / localhost:7687 (Bolt)

Python 3.11+ with a virtual environment:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

---

Running the Server

uvicorn app.main:app --reload --port 8000

Open http://localhost:8000 in your browser.

---

Usage

1. Set a learning objective — type it in the centre field and press Enter or the arrow button. This enables the composer.

2. Add source material — one or more of:

   • Paste text directly into the composer
   • Attach local files (.txt, .md, .pdf, .py, .js, .ts, .json, .csv, .html)
   • Add URLs via the link button

3. Submit — the agent fetches, chunks, and routes the content through the appropriate RAG strategy, then generates a curated bundle and a dynamic challenge.

4. Answer the challenge — write your response in the answer field and submit. The LLM judge scores it, provides feedback, and marks it passed (≥ 0.7) or not.

5. Improved Solution — five seconds after your score appears, a button reveals an LLM-generated improved version of your answer with annotated improvements.

---

Project Structure

app/
  main.py              FastAPI BFF — endpoints and session store
  graph.py             LangGraph workflow definition
  models.py            Pydantic models
  tools/
    retrieval.py       URL / file / text ingestion tools
    qdrant_rag.py      Standard RAG via Qdrant
    neo4j_rag.py       Graph RAG via Neo4j
    evaluators.py      Faithfulness + answer + improve evaluators
frontend/
  index.html
  style.css
  app.js
docker-compose.yml     Qdrant + Neo4j services
requirements.txt