研究论文AI对话助手

Quick Start

pip install deepxiv-sdk

On first use, deepxiv auto-registers a free anonymous token (1,000 requests/day) and saves it to ~/.env:

deepxiv search "agentic memory" --limit 5

For the full stack (MCP server + built-in research agent):

pip install "deepxiv-sdk[all]"

Examples

See examples/: quickstart.py, example_reader.py, example_agent.py, example_advanced.py, example_error_handling.py.

deepxiv-sdk

DeepXiv is an agent-first paper search and progressive reading tool.

Install it with pip, start using it immediately, and let the CLI auto-register an API token on first use — no setup needed before your first query.

### 🚦 Service status — live status page - 🟢 arXiv retrieval & reading — online. We aim for a T+1 sync with arXiv (subject to arXiv's own ~1-day API latency). - 🔴 bioRxiv / medRxiv — temporarily down due to a server-side issue. We're working to restore it as soon as possible. Related commands return 503 in the meantime. - 🔑 Lost your token? Recover it at data.rag.ac.cn/token-lookup (Google sign-in supported). - ℹ️ Data processing is currently trying a broader mix of models. If a TLDR looks off (e.g. truncated thinking content), please open an issue — we'll fix it.

• 🚦 Live Status: https://data.rag.ac.cn/status
• 📚 API Documentation: https://data.rag.ac.cn/api/docs
• 📄 Technical Report:
• 📖 中文文档: README.zh.md

<p align="center"> <img src="./assets/demo.gif" width="60%"> </p>

🚀 Live Demo: built on the deepxiv CLI in ~1 hour with vibe coding — try the DeepResearch demo. A full-stack research platform is on the way.

---

CLI Reference

Python SDK

```python from deepxiv_sdk import Reader

reader = Reader()

Unified retrieve endpoint; arXiv by default.

results = reader.search("agent memory", size=5) for paper in results["result"]: print(paper["arxiv_id"], paper["score"], paper["title"])

Other endpoints

web = reader.websearch("karpathy") sc_meta = reader.semantic_scholar("258001") ```

CLI Skill

mkdir -p $CODEX_HOME/skills
ln -s "$(pwd)/skills/deepxiv-cli" $CODEX_HOME/skills/deepxiv-cli

For frameworks without native skill support, load skills/deepxiv-cli/SKILL.md as operating instructions.

Agent Workflows

Two ready-to-use workflows ship as reusable skills:

Review recent hot papers → skills/deepxiv-trending-digest/SKILL.md

```bash deepxiv trending --days 7 --limit 30 --json

Agent Integration

DeepXiv works well inside Codex, Claude Code, OpenClaw, and similar agent runtimes.

Troubleshooting

• Do I need a token? No — some papers are free, and a token is auto-created on first use.
• Max search results? 100 per request; use --offset / offset= to paginate.
• A search returns 0 results? Loosen filters — stacked --date-* + --min-citations constraints can over-narrow the result set.
• Timeouts? The Reader retries (max 3) with exponential backoff. Customize with Reader(timeout=120, max_retries=5).
• Can I cache content? Yes — cache locally after fetching; paper content doesn't change.
• Which LLMs does the agent support? Any OpenAI-compatible API (OpenAI, DeepSeek, OpenRouter, local Ollama, …).
• Agent errors with Reasoning content is only supported as the last assistant message? Thinking/reasoning models (MiMo, DeepSeek-R1, …) need thinking disabled for multi-round tool use. Use deepxiv agent query "…" --disable-thinking, or in Python Agent(..., enable_thinking=False) (equivalently extra_body={"enable_thinking": False}).
• Agent keeps retrying a failing tool? When the data service is down, the agent now trips a circuit breaker after a few consecutive service failures and returns a best-effort answer instead of looping. Tune with Agent(..., max_consecutive_failures=N) (0 disables it).
• agent.add_paper() on a brand-new paper? It returns False (instead of raising) when the paper isn't found or isn't indexed yet — very recent papers (<1–3 days old) often aren't. Genuine errors (auth, rate limit, 5xx) still raise. To handle the exception directly: from deepxiv_sdk import NotFoundError (also available as from deepxiv_sdk.exceptions import NotFoundError).
• bioRxiv / medRxiv returns 503? Known outage — see the status page.