GraphLex AI

GraphLex AI is a local document assistant for PDF ingestion, retrieval, and question answering. It uses Qdrant for vector storage, Ollama for local model inference, and LangGraph for the agent workflow.

Features

• PDF ingestion with page-level chunking and metadata.
• Qdrant-backed similarity search with optional file filtering.
• LangGraph agent flow with retrieval, grading, rewriting, and generation.
• Streamlit interface for uploading documents and asking questions.
• Local model configuration with qwen2.5:3b.

Tech Stack

• Python 3.11
• Poetry
• LangChain and LangGraph
• Qdrant
• Ollama
• Streamlit

Project Layout

GraphLex/
├── docker-compose.yml
├── pyproject.toml
├── README.md
├── src/
│   ├── database.py
│   ├── engine.py
│   └── ui.py
└── tests/
    ├── test_agent_logic.py
    ├── test_agent_traces.py
    ├── test_final_answer.py
    └── test_multiple_files.py

Setup

Install dependencies with Poetry:

poetry install

Activate the environment if needed:

poetry shell

Local Services

Start Qdrant with Docker:

docker compose up -d qdrant

Make sure Ollama is running locally and that the qwen2.5:3b model is available.

Usage

Run the Streamlit UI:

poetry run streamlit run src/ui.py

The UI lets you upload a PDF, index it, and ask questions about the document.

Core Workflow

1. src/database.py loads PDFs, splits them into chunks, and stores them in Qdrant.
2. src/engine.py runs the LangGraph agent and handles retrieval, grading, query rewriting, and answer generation.
3. src/ui.py provides the Streamlit interface for upload and chat.

Test Scripts

The repository includes script-style checks under tests/:

• tests/test_final_answer.py runs a single end-to-end question.
• tests/test_agent_logic.py checks a positive and negative query.
• tests/test_agent_traces.py prints the final graph state.
• tests/test_multiple_files.py checks file-specific filtering.

Run them with Poetry:

poetry run python tests/test_final_answer.py
poetry run python tests/test_agent_logic.py
poetry run python tests/test_agent_traces.py
poetry run python tests/test_multiple_files.py

Model Configuration

The current default embedding and chat model is qwen2.5:3b via Ollama. If you change the model, update the comments and README so they match the code.

Notes

• Generated PDFs and Qdrant storage are intentionally ignored in Git.
• The repository uses dev for active implementation work.