Ptah: Building the Future of Research with Vision and Language

This repository contains the official implementation of our paper, "Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness for Interleaved Report Generation."

Authors: Chenghao Zhang, Guanting Dong, Yufan Liu, Tong Zhao, Xiaoxi Li, and Zhicheng Dou

Preparation

conda create -n ptah python=3.11
conda activate ptah
pip install -r requirements.txt

Required environment variables are loaded from .env when present:

SERPER_API_KEY=...
JINA_API_KEY=...
OPENAI_API_KEY=...

Jina Reader is called directly from the server through https://r.jina.ai. No local SSH tunnel or local Jina proxy is required.

Serve Local Models

Start Qwen3 on CUDA 0,1:

bash scripts/serve_llm.sh

Start Qwen3-VL on CUDA 2,3:

bash scripts/serve_vlm.sh

Defaults:

LLM: models/Qwen3-32B at http://localhost:8000/v1/
VLM: models/Qwen3-VL-32B-Instruct at http://localhost:8001/v1/

Run Reports

Run a custom report:

bash scripts/run_custom.sh

Run Deep Consult tasks:

bash scripts/run_dc.sh

Run DeepResearch Bench tasks:

bash scripts/run_drb.sh

Outputs

DC outputs:

outputs/dc/report_<id>.json
outputs/dc/report_<id>.html
.cache/dc/question_<id>/

DRB outputs:

outputs/drb/report_<id>.json
outputs/drb/report_<id>.html
.cache/drb/question_<id>/

Custom outputs:

outputs/custom/report.json
outputs/custom/report.html
.cache/custom/

Evaluation

DeepConsult

Run DC evaluation:

bash scripts/eval_dc.sh

Run DC PtahEval:

bash scripts/eval_dc_ptaheval.sh

DeepResearch Bench

Run DRB RACE:

bash scripts/eval_drb_race.sh

Run DRB FACT:

bash scripts/eval_drb_fact.sh

Run DRB PtahEval:

bash scripts/eval_drb_ptaheval.sh

Citation

@article{zhang2026ptah,
  author       = {Chenghao Zhang and
                  Guanting Dong and
                  Yufan Liu and
                  Tong Zhao and
                  Xiaoxi Li and
                  Zhicheng Dou},
  title        = {Towards Verifiable Multimodal Deep Research: A Multi-Agent Harness
                  for Interleaved Report Generation},
  journal      = {CoRR},
  volume       = {abs/2605.29861},
  year         = {2026},
  url          = {https://doi.org/10.48550/arXiv.2605.29861},
  doi          = {10.48550/ARXIV.2605.29861},
  eprinttype   = {arXiv},
  eprint       = {2605.29861},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2605-29861.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}