LATTICE turns retrieval into an LLM-driven navigation problem over a semantic scaffold for computational tractability needed for large corpora.
LATTICE proposes an LLM-native retrieval paradigm that combines the efficiency of hierarchical search with the reasoning power of modern large language models. Instead of relying on a static retriever + reranker pipeline or attempting to place a large corpus directly in an LLM context, LATTICE organizes the corpus into a semantic tree and uses an LLM as an active search agent that navigates that tree. This design yields logarithmic search complexity while preserving the LLM’s ability to perform nuanced, multi-step relevance judgments for complex, reasoning-heavy queries.
Read more in the blog / paper or try it out in the colab notebook
-
Clone the repository:
git clone https://github.com/nilesh2797/lattice cd lattice mkdir results trees -
Install dependencies:
pip install -r src/requirements.txt
-
Download pre-built semantic trees:
git clone https://huggingface.co/datasets/quicktensor/lattice-bright-trees ./trees/BRIGHT
-
Set up API credentials:
export GOOGLE_API_KEY=your_api_key_here
Run a single experiment:
cd src; python run.py --subset biology --tree_version bottom-up --num_iters 20Batch Experiments
cd src; bash run.sh| Parameter | Description | Default |
|---|---|---|
--subset |
Dataset subset (biology, economics, etc.) | Required |
--tree_version |
Tree construction method (bottom-up/top-down) | Required |
--num_iters |
Number of retrieval iterations | 20 |
--max_beam_size |
Beam size during traversal | 2 |
--relevance_chain_factor |
Weight for current score in path relevance | 0.5 |
--reasoning_in_traversal_prompt |
Enable reasoning (thinking budget) | -1 (enabled) |
--rerank |
Additional reranking on final results | False |
--load_existing |
Resume from checkpoint defined by hyperparams | False |
--suffix |
Experiment name suffix | - |
For a complete list, see src/hyperparams.py.
lattice/release/
├── src/
│ ├── run.py # Main execution script
│ ├── run.sh # Batch execution wrapper
| ├── run.ipynb # Jupyter notebook for running / debugging experiments
│ ├── hyperparams.py # Hyperparameter definitions
│ ├── tree_objects.py # Semantic tree and sample objects
│ ├── llm_apis.py # LLM API wrappers
│ ├── prompts.py # Prompt templates
│ ├── utils.py # Utility functions
│ └── calib_utils.py # Calibration utilities
├── trees/
│ └── BRIGHT/ # Pre-built semantic trees
├── results/
│ └── BRIGHT/ # Experiment results
└── logs/ # Execution logs
If you find this work helpful, please cite:
@article{gupta2025lattice,
title={LLM-Guided Hierarchical Retrieval},
author={Gupta, Nilesh and Chang, Wei-Cheng and Bui, Ngot and Hsieh, Cho-Jui and Dhillon, Inderjit S.},
journal={arXiv preprint arXiv:2510.13217},
year={2025}
}