Tongyi DeepResearch SOP

Appearance

Test ​

Status: NEEDS-ACCOUNT (GitHub clone + API key required) ​

Classification: NEEDS-ACCOUNT

Tongyi DeepResearch is not available as a published PyPI package. The project exists only on GitHub at github.com/Alibaba-NLP/DeepResearch. It requires a manual clone, install, and an Alibaba Cloud API key (or compatible LLM endpoint) before any inference is possible.

What is required ​

  1. Clone the repository:

    bash
    git clone https://github.com/Alibaba-NLP/DeepResearch.git
cd DeepResearch

  2. Create and activate the conda environment:

    bash
    conda create -n react_infer_env python=3.10
conda activate react_infer_env
pip install -r requirements.txt

  3. Provide API keys for:

    • Alibaba Cloud (or an OpenAI-compatible LLM endpoint)
    • Serper (web search)
    • Jina (reader/scraper)

Step 1: Verify Python environment ​

bash
conda activate react_infer_env
python --version

Expected output:

Python 3.10.0

Any version other than 3.10.0 may cause dependency conflicts.

Step 2: Verify dependencies installed ​

bash
python -c "import vllm; import openai; import dotenv; print('deps OK')"

Expected output:

deps OK

If any import fails, re-run pip install -r requirements.txt inside the activated environment.

Step 3: Validate your question file format ​

Create a minimal test file:

bash
echo '{"question": "What is the capital of France?", "answer": "Paris"}' > eval_data/smoke_test.jsonl

Validate it parses correctly:

bash
python -c "
import json
with open('eval_data/smoke_test.jsonl') as f:
    for line in f:
        obj = json.loads(line)
        assert 'question' in obj and 'answer' in obj
        print('Format OK:', obj['question'])
"

Expected output:

Format OK: What is the capital of France?

Step 4: Dry-run config check ​

Set MODEL_PATH, DATASET, and OUTPUT_PATH in run_react_infer.sh, then confirm the script is executable and the paths exist:

bash
bash -n run_react_infer.sh && echo "Script syntax OK"
ls eval_data/smoke_test.jsonl && echo "Dataset file found"
mkdir -p outputs && echo "Output dir ready"

Expected output:

Script syntax OK
eval_data/smoke_test.jsonl
Dataset file found
Output dir ready

Step 5: Full end-to-end run (with GPU or OpenRouter) ​

Once Steps 1-4 pass, run inference on the smoke test file:

bash
# In run_react_infer.sh, temporarily set:
# DATASET="eval_data/smoke_test.jsonl"
# OUTPUT_PATH="./outputs/smoke"

bash run_react_infer.sh

Success looks like: A JSON file appears in ./outputs/smoke/ containing the model's answer to "What is the capital of France?". The agent should issue one or two tool calls (web search or direct reasoning) and return Paris as the final answer within a few tool-use steps.

Notes ​

  • This smoke test was not run locally in this SOP session because the full stack requires a downloaded 30B model (significant disk and GPU VRAM), active API keys for Serper, Jina, and at minimum one OpenAI-compatible endpoint, and an Alibaba Cloud account for the Tongyi model.
  • Steps 1-4 above can be run without GPU or API keys and are sufficient to confirm the environment is ready for a real inference job.
  • tested=false reflects this. There is no published pip package and no one-liner install available.