Understand the project
Practical overview
When it fits
Use it when you want research code for retrieval-guided query generation or a documented local workflow for model-based query rewriting.
Who it is for
- Researchers studying reinforcement learning for search and information retrieval.
- Developers prototyping search agents or locally hosted query-rewriting services.
Documented uses
- Train a language model to iteratively generate search queries using retrieval results as feedback.
- Experiment with retrieval-based rewards for search tasks.
- Host the model and rewrite an input query using the repository’s documented script.
Before you choose
Requirements & limits
The project is public source code under the MIT license. Its documented query-rewriting workflow hosts a model locally.
- The training workflow uses retrieval metrics as rewards rather than labeled query-rewrite pairs.
- General setup documents Python 3.9, PyTorch, vLLM, Ray and the project’s verl code.
- Sparse or dense retrieval may need additional tools such as Pyserini, FAISS and Java; search-engine retrieval requires an API key.
- The repository documents query rewriting through a locally hosted vLLM model and Python script.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Source-based practical guide
Set up the documented query-rewriting workflow
Host a model with vLLM and submit a query to Deepretrieval’s query-rewriting script.
Before you start
- A Python development environment with the repository’s documented general dependencies, including vLLM, Ray and the verl code.
- A model and local port selected for the vLLM host.
Steps
Install the documented general dependencies in the repository’s Python 3.9 Conda environment.
Run the command below, specifying the local port and model as needed.
sh vllm_host.shRun the command below.
python query_rewrite.py --query "Who built DeepRetrieval in 2025?"
Expected outcome
The documented workflow returns an original query and a rewritten query.
Guide limitations
- This describes a locally hosted workflow, not a public hosted API endpoint.
- The README does not specify model hardware requirements in this API example.
Evidence and transparency
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 15 October 2026
- Original website title
- pat-jj/DeepRetrieval · Official homepage title · retained in its source language