How mohithpeta/graph-based-rag works
AGRAG is a fashion recommendation system that turns image or text queries into recommendations. Its principal runtime flow is query embedding, Neo4j vector retrieval with optional category and gender constraints, then reranking and diversification. The README also documents a model-training and catalog-ingestion path that prepares the embeddings and knowledge graph. Evaluation scripts measure retrieval quality and diversity. Graph-neighbor expansion is a documented capability, but its implementation was not visible in the sampled source, so that relationship is identified as documented rather than inferred from code.
Main entry points: agent.py AGRAG agent · retriever.py Vector retriever · build_embeddings.py Embedding builder · evaluate_retrieval.py Retrieval evaluation
Components 18
Catalog preparation
- FashionGen data
- Embedding builder
build_embeddings.py - Vectors and metadata
build_embeddings.py - Knowledge graph loader
kg/load_to_neo4j.py - Neo4j catalog
Recommendation runtime
- Query embedder
agrag/embedder.py - AGRAG agent
agrag/agent.py - Vector retriever
agrag/retriever.py - Graph expansion
agrag - Hybrid reranking
agrag/agent.py - MMR diversification
agrag/mmr.py - Recommendations
- Search interface
agrag/run_agrag.py
Model training
- CLIP training
src/training/trainer.py - Fine-tuned CLIP
Evaluation
- Retrieval evaluation
evaluate_retrieval.py - Ranking metrics
agrag/eval/metrics.py
Other components
- User query
Connections 19
- CLIP training → Fine-tuned CLIP: saves model
src/training/trainer.py - Fine-tuned CLIP → Embedding builder: loads model
build_embeddings.py - FashionGen data → Embedding builder: embeds catalog
build_embeddings.py - Embedding builder → Vectors and metadata: writes vectors
build_embeddings.py - Vectors and metadata → Knowledge graph loader: loads data
kg/load_to_neo4j.py - Knowledge graph loader → Neo4j catalog: writes items
kg/load_to_neo4j.py - User query → Search interface: enters query
agrag/run_agrag.py - Search interface → AGRAG agent: calls search
agrag/run_agrag.py - AGRAG agent → Query embedder: embeds query
agrag/agent.py - Query embedder → Fine-tuned CLIP: loads model
agrag/embedder.py - AGRAG agent → Vector retriever: retrieves items
agrag/agent.py - Vector retriever → Neo4j catalog: queries vectors
agrag/retriever.py - AGRAG agent → Hybrid reranking: dispatches rerank
agrag/agent.py - Hybrid reranking → MMR diversification: diversifies
agrag/agent.py - AGRAG agent → Graph expansion: uses traversal
README.md - MMR diversification → Recommendations: selects results
- Vector retriever → Recommendations: returns items
- Retrieval evaluation → AGRAG agent: evaluates search
evaluate_retrieval.py - Retrieval evaluation → Ranking metrics: computes ranking
evaluate_retrieval.py