Graph-Grounded Temporal RAG

Contradiction-Resilient Question Answering Over Evolving Documents

Bachelor Thesis — Mekelle Institute of Technology, Mekelle University (2026)

The Problem

Standard RAG systems are temporally blind. When a document has been amended multiple times, they retrieve all versions equally and produce contradictory or outdated answers. This system fixes that.

How It Works

1. Parse — extracts text from PDF/DOCX with metadata headers
2. Chunk — splits into 1,200-char overlapping segments (3,408 chunks)
3. Embed — converts chunks to 384-dim vectors (all-MiniLM-L6-v2)
4. Graph — Neo4j encodes SUPERSEDES relationships between versions
5. Retrieve — hybrid search (vector + BM25 + keyword) fused via RRF
6. Rerank — cross-encoder (ms-marco-MiniLM-L-6-v2) rescores top 30
7. Answer — Llama 3.2:3B generates grounded answer locally via Ollama

Architecture

PDF/DOCX → Parser → Chunker → Embedder → ChromaDB → Entity Extractor → Neo4j Graph User Question → Hybrid Search → Graph Resolve → Reranker → Llama 3.2 → Answer

Tech Stack

Component	Technology
LLM	Llama 3.2:3B (Ollama)
Embeddings	all-MiniLM-L6-v2
Reranker	ms-marco-MiniLM-L-6-v2
Vector DB	ChromaDB
Graph DB	Neo4j
Sparse search	BM25Okapi (rank-bm25)
PDF parsing	PyMuPDF
Web UI	Streamlit
REST API	FastAPI

Setup

Requirements

• Python 3.11
• Neo4j Desktop (running locally)
• Ollama with llama3.2:3b pulled

Installation

git clone https://github.com/WeleTeklay/Graph-Grounded-Temporal-RAG
cd Graph-Grounded-Temporal-RAG
python -m venv venv
venv\Scripts\activate        # Windows
pip install -r requirements.txt
python -m spacy download en_core_web_sm

Configure

Copy .env.example to .env and fill in your Neo4j password: NEO4J_URI=bolt://localhost:7687 NEO4J_USER=neo4j NEO4J_PASSWORD=your_password OLLAMA_MODEL=llama3.2:3b

Run the pipeline

python src/01_parse_pdfs.py
python src/02_chunk_documents.py
python src/04_build_graph.py
python src/05_create_index.py

Launch

# Web UI
streamlit run app.py

# REST API
python api.py

Results

• 98% diagnostic pass rate on vector database, hybrid retrieval, reranking, and LLM integration
• Average query response time: 3,481 ms on standard laptop
• Fully offline — no API keys, no internet required
• Works on any evolving document corpus (legal, medical, financial, HR)

Authors

• Weldesemayat Teklay Gebre — github.com/WeleTeklay
• Gebregergs Mekonen — github.com/gere047

Advisors

Assefa Tesfay (PhD) and Yaecob Girmay (MSc.) — Mekelle Institute of Technology - Mekelle University