This directory provides a PASA service layout that can be called directly by DeepReviewer:

• Two vLLM OpenAI-compatible inference services (crawler + selector)
• One Flask orchestrator (pasa_server.py)
• One unified start/stop script (start_pasa_server.sh)

DeepReviewer accesses PASA with:

• PAPER_SEARCH_BASE_URL=http://127.0.0.1:8001
• PAPER_SEARCH_ENDPOINT=/pasa/search

• PASA official repository: https://github.com/bytedance/pasa
• Official README: https://github.com/bytedance/pasa/blob/main/README.md
• Crawler model: https://huggingface.co/bytedance-research/pasa-7b-crawler
• Selector model: https://huggingface.co/bytedance-research/pasa-7b-selector
• Dataset: https://huggingface.co/datasets/CarlanLark/pasa-dataset
• Serper API signup: https://serper.dev/

• Linux + NVIDIA GPU
• Python 3.10+ (3.11 recommended)
• Working CUDA environment (compatible with your PyTorch / vLLM versions)
• Network access to Hugging Face, arXiv, and Serper

Run this in your Python environment:

cd <repo_root>/pasa
pip install --upgrade pip
pip install \
  torch transformers \
  vllm "openai>=1.52,<1.76" \
  flask flask-cors \
  requests httpx arxiv \
  beautifulsoup4 lxml

Notes:

• start_pasa_server.sh checks import vllm before startup.
• pasa/pasa/utils.py loads the local paper DB at import time, so configure these file paths first.

Example using huggingface-cli:

# crawler
huggingface-cli download bytedance-research/pasa-7b-crawler \
  --local-dir /data/models/pasa-7b-crawler

# selector
huggingface-cli download bytedance-research/pasa-7b-selector \
  --local-dir /data/models/pasa-7b-selector

Then point paths in .pasa_env to your local model directories.

The current code reads:

• PASA_PAPER_DB (for example: cs_paper_2nd.zip)
• PASA_PAPER_ID (for example: id2paper.json)

Download them from the official dataset page, store locally, then configure them in .pasa_env.

Recommended setup:

cd <repo_root>/pasa
cp .pasa_env.example .pasa_env.local
vim .pasa_env.local

Notes:

• pasa_server.py loads env files in this order: $PASA_ENV_FILE -> .pasa_env.local -> .pasa_env
• Keep machine-specific settings in .pasa_env.local and do not commit them

Key configuration example:

# GPU
PASA_GPU_ID=1

# Flask server
PASA_SERVER_HOST=0.0.0.0
PASA_SERVER_PORT=8001

# Model paths (must exist)
PASA_CRAWLER_PATH=/data/models/pasa-7b-crawler
PASA_SELECTOR_PATH=/data/models/pasa-7b-selector
PASA_PROMPTS_PATH=pasa/agent_prompt.json

# vLLM service endpoints
PASA_VLLM_HOST=127.0.0.1
PASA_VLLM_CRAWLER_PORT=8101
PASA_VLLM_SELECTOR_PORT=8102
PASA_VLLM_CRAWLER_URL=http://127.0.0.1:8101/v1
PASA_VLLM_SELECTOR_URL=http://127.0.0.1:8102/v1
PASA_VLLM_CRAWLER_MODEL_NAME=pasa-crawler
PASA_VLLM_SELECTOR_MODEL_NAME=pasa-selector

# Serper key (now read from env var; no longer hardcoded)
PASA_SERPER_API_KEY=your_serper_api_key
PASA_SERPER_SEARCH_URL=https://google.serper.dev/search

# Local paper DB (set real paths)
PASA_PAPER_DB=/data/pasa/cs_paper_2nd.zip
PASA_PAPER_ID=/data/pasa/id2paper.json

cd <repo_root>/pasa
bash start_pasa_server.sh

cd <repo_root>/pasa
bash start_pasa_server.sh --background

cd <repo_root>/pasa
bash start_pasa_server.sh --stop

cd <repo_root>/pasa
bash start_pasa_server.sh --restart

curl http://127.0.0.1:8001/health

Expected fields:

• "status": "healthy"
• "crawler_ready": true
• "selector_ready": true

curl -X POST http://127.0.0.1:8001/pasa/search \
  -H "Content-Type: application/json" \
  -d '{
    "query": "Papers about contrastive learning",
    "expand_layers": 1,
    "search_queries": 2,
    "search_papers": 5,
    "expand_papers": 5,
    "threads_num": 0
  }'

cd <repo_root>/pasa
python test_pasa_decoupling.py

• GET /
• GET /health
• POST /pasa/search
• POST /pasa/search_async
• GET /pasa/jobs/<job_id>
• GET /pasa/jobs/<job_id>/result
• DELETE /pasa/jobs/<job_id>

Set the following in <repo_root>/.env:

PAPER_SEARCH_BASE_URL=http://127.0.0.1:8001
PAPER_SEARCH_ENDPOINT=/pasa/search
PAPER_SEARCH_API_KEY=

1. vllm import fails

• Ensure the Python environment used by the startup script has vllm installed.

2. Model path does not exist

• Check PASA_CRAWLER_PATH and PASA_SELECTOR_PATH in .pasa_env.

3. /health is unhealthy

• Ensure PASA_VLLM_*_MODEL_NAME matches the vLLM --served-model-name.
• Check logs: /tmp/pasa_vllm_crawler.log, /tmp/pasa_vllm_selector.log, /tmp/pasa_server.log.

4. /pasa/search errors or returns empty results

• Ensure PASA_SERPER_API_KEY is configured correctly.
• Ensure your network/proxy can access google.serper.dev and arXiv.

5. Importing pasa/pasa/utils.py fails on startup

• Usually caused by invalid PASA_PAPER_DB or PASA_PAPER_ID paths. Fix paths and restart.