---
license: mit
task_categories:
- text-generation
- text-retrieval
- text-ranking
tags:
- RAG
- Workload
- Real-World
configs:
- config_name: default
  data_files:
  - split: train
    path: "0_trace.jsonl"
---
<p align="center">
  <img src="imgs/logo.webp.png" width="40%" alt="logo">
</p>

<h1 align="center" style="font-size:42px; font-weight:700;">
  RAGPulse: A Real-World RAG Workload Trace to Optimize RAG Serving Systems
</h1>

<p align="center">
🌐 <a href="https://github.com/flashserve/RAGPulse">Github Link</a> |
🤗 <a href="https://github.com/flashserve/RAGPulse/tree/main/data">Workload Trace</a> |
📑 <a href="https://arxiv.org/pdf/2511.12979">Arxiv Paper</a> |
🤖 <a href="https://github.com/flashserve/RAGPulse/tree/main/example/single_online_instance">How to use?</a>
</p>

RAGPulse is a real-world RAG workload trace collected from an university-wide Q&A service scenario. The system has been serving over 40,000 students and faculties since April 2024, providing intelligent policy Q&A services. The trace contains a total of 7,106 records entries, sampled from one week of our Q&A service.

## Dataset Details

**There are currently 6 files in [data](https://github.com/flashserve/RAGPulse/tree/main/data)**

- `0_trace.jsonl` contains the key information of our trace. Totally 7,106 lines.
- `1_sys_prompt.jsonl` contains the hash_ids corresponding to the system prompt content and the length of the token corresponding to the hash_ids. Totally 6,898 lines.
- `2_passages.jsonl` contains the hash_ids corresponding to vector database, which holds a total of 8,302 lines of passages content, and the length of the token corresponding to the hash_ids.
- `3_history.jsonl` contains the hash_ids corresponding to user's 4,910 lines chat history content, and the length of the token corresponding to the hash_ids.
- `4_user_input.jsonl` contains the hash_ids corresponding to the question entered by the user content, which holds a total of 6,292 chunk hash_ids, and the length of the token corresponding to the hash_ids.
- `5_web_search.jsonl` contains the hash_ids corresponding to content retrieved online content, and the length of the token corresponding to the hash_ids.

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

```json
{
    "timestamp": "27", 
    "input_length": 3861, 
    "output_length": 127, 
    "hash_ids": {
        "sys_prompt": [8325, 8326, 11575], 
        "passages_ids": [6123, 7239, 6124, 1167, 7250, 5448], 
        "history": [15215], 
        "web_search": [20319, 20320], 
        "user_input": [23648]
        }, 
    "session_id": "1758081660427-xa8rbsd2uco1"
}
```

The above shows **one example** from **RAGPulse**. All textual content has been replaced with remapped hash IDs for user privacy. The detailed structure of the trace is as follows:

- `timestamp`: The request submission time,in seconds, measured from the beginning of the trace (12:00:00).
- `input_length`: Total token length of the request, including the system prompt, retrieved passages, chat history, web search, user input, and so on.
- `output_length`: Total token length of the output.
- `hash_ids`: A comprehensive collection of hash identifiers representing every component of the request's input. This includes unique IDs for the system prompt, all retrieved documents, user chat history, external web search results,and user's question.
- `session_id`: The conversation identifier to which the request belongs.

## Significance

### Why we need a RAG-Specific trace?

- **Multi-Stage Pipeline Complexity**<br>Unlike pure LLM inference traces, RAG involves retrieval, reranking, and generation stages. The latency, resource usage, and caching patterns are jointly influenced by these multi-step interactions — something existing LLM traces cannot capture.
- **Knowledge-Dependent Workload Dynamics**<br>RAG performance depends heavily on the underlying vector database and query patterns. Query similarity, embedding cache hit rate, and retrieval latency all vary with real-world user behavior — which must be reflected in a dedicated trace.
- **Cross-Request Context Dependency**<br>In many RAG services (e.g., campus QA or document understanding), consecutive queries are often semantically related or share similar knowledge needs. Capturing this inter-request correlation is critical for optimizing KV cache reuse, batching, and scheduling.
- **Retrieval-Serving Co-Optimization Need**<br>A RAG trace enables joint optimization of both retrieval and generation components, rather than treating them as independent black boxes. This helps design better serving systems, such as dynamic retrieval caching or query-adaptive batching.
- **Bridging Research and Real-World Deployment**<br>Existing LLM traces are mostly from synthetic or general-purpose chat workloads. A RAG-specific trace provides grounded insights for RAG system design, cost-performance tradeoffs, and scheduling strategies in real deployments.
</details>

### What can we do with this RAG trace?

- **System Performance Analysis**
Study end-to-end latency breakdown across retrieval, ranking, and generation stages.
- **Scheduling and Caching Optimization**
Evaluate KV cache reuse, adaptive batching, and retrieval caching under real correlation patterns.
- **Retrieval Quality and Cost Tradeoff**
Explore how retrieval depth, embedding model, and DB latency affect overall service cost and quality.
- **Workload Modeling and Simulation**
Build realistic RAG simulators or benchmarks for academic or industrial RAG system evaluation.
- **Agent Evaluation Studies**
Track how real-world RAG workloads evolve toward agent-style reasoning and multi-turn task-solving.

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->
Please kindly cite our paper if you find the paper or the trace is useful:

```bibtex
@misc{RAGPulse,
      title={RAGPulse: An Open-Source RAG Workload Trace to Optimize RAG Serving Systems}, 
      author={Zhengchao Wang and Yitao Hu and Jianing Ye and Zhuxuan Chang and Jiazheng Yu and Youpeng Deng and Keqiu Li},
      year={2025},
      eprint={2511.12979},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2511.12979}, 
}
```