Graphazon

Graphazon is an applied research-focused project that combines query understanding, knowledge graph construction, and learning-to-rank models for Amazon-style product search. The system is powered byLLMs · Knowledge Graph Extraction · Agentic AI · Recommender Systems.

---

Motivation

Product search is a complex problem that requires:

• Understanding user intent from ambiguous queries using LLM reasoning (LangChain)
• Mapping queries to relevant products, including substitutes and complements
• Leveraging structured and unstructured product information via a Knowledge Graph
• Improving ranking with graph-derived and semantic features

Graphazon addresses this by:

1. Using LangChain for query understanding, intent detection, and attribute extraction.
2. Building a task-specific Knowledge Graph from metadata and product text, including structured and LLM-assisted triples.
3. Incorporating graph-derived features into a learning-to-rank model trained on real-world data.

---

Technologies & Features

• LangChain: LLM chains for query rewriting, intent detection, and attribute extraction.
• Knowledge Graphs (KGs): Structured + LLM-assisted KG construction to capture product relationships (Exact, Substitute, Complement).
• Embeddings & Features: Combine KG, product metadata, and query embeddings for ranking.
• Learning-to-Rank: Neural / gradient-boosted ranking models trained on ESCI labels.
• Pipelines: Modular orchestration for end-to-end experiments.

---

Dataset

Graphazon uses the Amazon Shopping Queries Dataset (ESCI Benchmark):

• Provides queries, candidate products, and ESCI labels (Exact / Substitute / Complement / Irrelevant)
• Includes multilingual queries (English, Spanish, Japanese)
• Available as a git submodule in data/shopping-queries-dataset/raw/amazon-science/esci-data/

Directory structure:

data/
└── shopping-queries-dataset/
├── raw/
│ └── amazon-science/esci-data/ # Git submodule from amazon-science
└── processed/ # cleaned queries, products, KG triples, splits

To initialize the dataset:

git submodule update --init --recursive

Project Structure

Graphazon/
├── data/                  # datasets
├── src/                   # source code
│   ├── query_understanding/
│   ├── knowledge_graph/
│   ├── ranking/
│   ├── evaluation/
│   └── pipelines/
├── experiments/           
├── notebooks/             # EDA, visualization, shows usage
├── scripts/               # utility scripts
├── pixi.toml              # environment definition
└── README.md

• query_understanding/: LLM chains and prompts for query parsing, intent detection, attribute extraction

• knowledge_graph/: KG schema, structured and LLM-assisted extractors, graph feature generation

• ranking/: Learning-to-rank models using features from KG and embeddings

• evaluation/: Metrics, ablations, error analysis

• pipelines/: Orchestration scripts that run end-to-end processes

Installation

Graphazon uses pixi for environment management.

pixi install

Then initialize the ESCI dataset:

git submodule update --init --recursive

Running Experiments

# Run query understanding
pixi run python -m src.pipelines.run_query_understanding

# Build the knowledge graph
pixi run python -m src.pipelines.build_kg

# Combine the features
pixi run python -m src.pipelines.build_features

# Train the learning-to-rank model
pixi run python -m src.pipelines.train_ranker

# Evaluate results
pixi run python -m src.pipelines.evaluate

Find examples with command-line arguments in scripts/

License

This project is for research and educational purposes. ESCI dataset is provided under its own license.