title	Adaptive Study Agent
emoji	📚
colorFrom	blue
colorTo	purple
sdk	gradio
sdk_version	6.9.0
app_file	app.py
pinned	false
private	true

Adaptive Study Agent

A two-model LLM self-examination simulation built with LangGraph, Claude (Anthropic), and GPT-4o-mini (OpenAI). The agent reads any document you provide and runs a fully autonomous study loop — no human answers anything.

How the two models collaborate:

• GPT-4o-mini generates comprehension questions from document chunks (temperature 0.7 — creative) and evaluates the answers against the source material (temperature 0.0 — deterministic)
• Claude Sonnet answers the questions using RAG retrieval from ChromaDB (temperature 0.3 — balanced)
• OpenAI text-embedding-3-small handles document chunking and embedding into ChromaDB only — not used for reasoning

The purpose is to probe where Claude's understanding of the document breaks down — GPT acts as the examiner, Claude as the student. When Claude scores below the mastery threshold, the agent re-reads the weak chunk and tries again.

The output is a structured session report revealing Claude's weak areas within your document — useful for identifying conceptually dense or underrepresented sections in any text.

Applicable to any domain — ML papers, medical literature, legal documents, textbooks — anything in PDF or TXT format.

---

Research Connection

This is a standalone extended example project inspired by ongoing research on multi-agent knowledge systems. The core idea — using retrieval-augmented self-evaluation to surface knowledge gaps — is the single-agent version of a feedback mechanism explored at scale in that research. There is no shared infrastructure or data pipeline between the two.

---

---

Architecture

The agent operates as a LangGraph state machine with conditional branching. After evaluating each answer, the agent decides whether to re-read weak material, continue to the next question, or finalize the session.

---

Tech Stack

Component	Technology	Purpose
Agent framework	LangGraph	Stateful loops with conditional branching
Examiner LLM	GPT-4o-mini (OpenAI)	Question generation (0.7) + evaluation (0.0)
Student LLM	Claude Sonnet 4 (Anthropic)	Answering questions via RAG (0.3)
Embeddings	OpenAI text-embedding-3-small	Document chunking and embedding into ChromaDB only
Vector store	ChromaDB (local, embedded)	No Docker required
Document parsing	PyMuPDF (fitz)	PDF support
UI	Gradio	Web interface and Hugging Face Spaces deploy
Package manager	uv	Dependency management

---

Setup

1. Install dependencies

uv sync

2. Configure environment variables

Create a .env file in the project root:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...

The Anthropic key powers the LLM (question generation, answering, evaluation). The OpenAI key is used only for embeddings.

3. Add documents

Place PDF or TXT files in data/documents/.

---

Usage

Command line

# Run with a document
uv run python src/main.py --doc data/documents/attention_is_all_you_need.pdf

# Override the mastery threshold (default: 0.75)
uv run python src/main.py --doc data/documents/myfile.pdf --threshold 0.8

# Persist the ChromaDB collection between runs
uv run python src/main.py --doc data/documents/myfile.pdf --persist

Gradio web interface

uv run python app.py

The web interface allows you to upload a document, configure the mastery threshold, start a study session, and view the resulting session report from the browser.

Running tests

uv run pytest tests/ -v

---

Configuration

The following constants in src/graph/edges.py control the study loop:

Parameter	Default	Description
MASTERY_THRESHOLD	0.75	Score needed to skip re-read
MIN_QUESTIONS	10	Minimum questions before mastery check
MAX_REREAD_CYCLES	3	Max re-read attempts per weak chunk

The mastery threshold can also be overridden at runtime via the --threshold flag or the Gradio slider.

---

Output Format

Each session produces a Markdown report in output/session_reports/:

# Study Session Report
Date: 2026-03-16
Document: attention_is_all_you_need.pdf

## Summary
- Questions asked: 14
- Questions correct (score >= 0.75): 11
- Final mastery score: 0.81
- Re-read cycles triggered: 3

## Weak Areas
- Multi-head attention computation
- Positional encoding formula

## Q&A Log
### Q1
Question: What is the purpose of the scaling factor in dot-product attention?
Answer: ...
Score: 0.9
...

---

Author

Halima Akhter PhD Student in Computer Science Specialization: ML, Deep Learning, Bioinformatics GitHub: github.com/Mituvinci