This repository contains a prototype implementation of VersionRAG (Versioned Document Retrieval-Augmented Generation) — a retrieval-augmented question answering system designed to handle versioned document collections. The project was developed for academic purposes as part of a bachelor's thesis.

⚠️ Prototype Notice
This project is a research prototype developed for academic evaluation purposes only.
It is not optimized for production use, and the implementation prioritizes clarity and experimentation over code quality or performance.

VersionRAG is a retrieval-augmented generation (RAG) pipeline developed specifically for working with versioned document collections. It supports version-aware retrieval and reasoning by explicitly modeling:

• document and version metadata,
• changes across document states (e.g., additions, deprecations),
• and a version-sensitive graph structure.

The core idea is to build a version graph during indexing, capturing relationships between document versions and tracking their evolution. During retrieval, VersionRAG uses a hybrid strategy: depending on the query type, it either leverages the graph structure directly (e.g., for change reasoning), or performs vector-based content retrieval with version constraints applied via the graph.

In contrast to standard RAG or GraphRAG, VersionRAG is designed to maintain temporal alignment and version specificity, making it suitable for scenarios that require high precision across evolving documentation.

For benchmarking purposes, the repository also includes independent implementations of:

• Naive RAG (baseline with standard dense retrieval),
• GraphRAG (retrieval via knowledge graph).

These implementations are included solely for performance comparison and are not part of the VersionRAG system.

Figure: High-level architecture of the VersionRAG system. The indexing phase builds a version-aware graph; the retrieval phase selects the appropriate retrieval strategy based on the question context.

Figure: Conceptual structure of the version-aware document graph. Nodes represent categories, documents, versions, content, and tracked changes.

• Multi-document support
  Handles multiple versioned documents across different categories, automatically distinguishing between standard documents and changelogs. Indexing is fully automated with no manual intervention required.

• Version-aware graph construction
  Builds a structured graph representing documents, versions, and occured changes between subsequent versions.

• Change tracking
  Detects and links additions, removals, and deprecations across document versions.

• Version metadata indexing
  Extracts and stores version-specific metadata for targeted retrieval.

• Targeted retrieval capabilities

  • Retrieve content for a specific document version
  • List available document versions
  • Identify tracked changes between two versions

• LLM integration
  Supports dynamic configuration of the underlying language model used for generation.

• Baseline implementations included
  Includes reference implementations of standard (naive) RAG and GraphRAG for performance comparison.

Install dependencies listed in src/requirements.txt:

pip install -r src/requirements.txt

Switch into the src/ directory and run the main program:

cd src
python main.py

To run the system, a .env file must be placed in the src/ directory to provide necessary credentials for external services.

• OPENAI_API_KEY – for OpenAI-based language models
• GROQ_API_KEY – for Groq-based language models

⚙️ The selection of which LLMs are used is configured in src/util/constants.py.

• NEO4J_URI
• NEO4J_USER
• NEO4J_PASSWORD

• NEO4J_URI_AURA
• NEO4J_USERNAME_AURA
• NEO4J_PASSWORD_AURA
• AURA_INSTANCEID

⚠️ Make sure to never commit your .env file to version control.