deepxiv-sdk

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

🚀 Retrieval service upgraded with better accuracy and richer filters (source, authors, orgs, categories, advanced date ranges). Please upgrade to the latest deepxiv-sdk to pick up the new parameters: pip install -U deepxiv-sdk. See Search API changes (2026-04) for migration notes.

Install it with pip, start using it immediately, and let the CLI auto-register an API token on first use. No extra setup is required before your first query.

• 📚 API Documentation: https://data.rag.ac.cn/api/docs
• 🎥 Demo Video:
• 📄 Technical Report:
• 📖 中文文档: README.zh.md

---

🚀 Live Demo: I used vibe coding, based on deepxiv CLI, to build a DeepResearch demo in 1 hour — feel free to try it out! A full-stack research platform is on the way. Stay tuned.

---

What DeepXiv Does

DeepXiv is built around two core workflows that matter for agents:

1. Search + Progressive Content Access
2. Trending + Popularity signals

Instead of blindly loading full papers, DeepXiv lets agents read in layers, based on token budget and task value.

Quick Start

pip install deepxiv-sdk

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

deepxiv search "agentic memory" --limit 5

If you want the full stack including MCP and the built-in research agent:

pip install "deepxiv-sdk[all]"

CLI-First Workflow

The CLI is the primary interface. DeepXiv is designed so agents can work like researchers: search first, judge quickly, then read only the most valuable parts.

deepxiv search "agentic memory" --limit 5
deepxiv paper 2603.21489 --brief
deepxiv paper 2603.21489 --head
deepxiv paper 2603.21489 --section Analysis

Three commands matter most for progressive reading:

• --brief: decide whether a paper is worth deeper reading
• --head: inspect structure, sections, and token distribution
• --section: read only the most valuable parts such as Introduction, Method, or Experiments

This is the core DeepXiv idea: agents should not load full papers unless they truly need them.

CLI Features

1. Paper Search and Reading

Basic usage:

deepxiv search "transformer" --limit 10
deepxiv paper 2409.05591 --brief
deepxiv paper 2409.05591 --head
deepxiv paper 2409.05591 --section Introduction
deepxiv paper 2409.05591

New search examples (2026-04)

The unified retrieve endpoint supports rich filtering and three sources in one command. A few patterns worth knowing:

Switch source with one flag

deepxiv search "image generation on GenEval" --limit 5
deepxiv search "de novo protein design" --biorxiv --limit 5
deepxiv search "multimodal Alzheimer diagnosis" --medrxiv --limit 5

Filter by authors, orgs, and categories (comma-separated lists)

deepxiv search "image generation" \
  --authors "Shitao Xiao,Zheng Liu" \
  --orgs "Beijing Academy of Artificial Intelligence" \
  --categories cs.CV \
  --limit 5

--authors and --orgs are filters and ranking signals. --categories is a pure filter.

Convenience date filters (auto-mapped to the new date semantics)

# Papers from June 2025 with at least 50 citations
deepxiv search "image generation outperforming SDXL" \
  --date-from 2025-06 --date-to 2025-06 \
  --min-citations 50 --limit 3

# Anything after a given date
deepxiv search "agentic memory" --date-from 2026-03-01 --limit 20

Advanced date filter (exact / after / before / between)

# exact month
deepxiv search "image generation" \
  --date-search-type exact --date-str 2025-06 --limit 5

# between (pass --date-str twice)
deepxiv search "image generation" \
  --date-search-type between \
  --date-str 2025-06-01 --date-str 2025-07-01 \
  --limit 5

Pagination

deepxiv search "LLM alignment" --limit 10 --offset 0
deepxiv search "LLM alignment" --limit 10 --offset 10

Opt-in to fine reranking (off by default)

# Default: fast recall, no rerank
deepxiv search "transformer model" --limit 10

# Enable upstream fine rerank when you need better ordering
deepxiv search "transformer model" --use-fine-rerank --limit 10

JSON output for programmatic consumption

deepxiv search "agentic memory" --limit 20 --format json

The JSON payload follows the new response shape ({status, total_count, result: [...]}). See Search API changes (2026-04) for details.

Python equivalents

from deepxiv_sdk import Reader
reader = Reader()

# Source + filters + advanced date
hits = reader.search(
    "image generation outperforming SDXL on GenEval",
    source="arxiv",
    size=5,
    categories=["cs.CV"],
    authors=["Shitao Xiao", "Zheng Liu"],
    orgs=["Beijing Academy of Artificial Intelligence"],
    min_citation=50,
    date_search_type="exact",
    date_str="2025-06",
)
for paper in hits["result"]:
    print(paper["arxiv_id"], paper["score"], paper["title"])

# Date range the shortcut way
reader.search("LLM alignment", date_from="2026-03-01", size=20)

# Switch source without a separate method
reader.search("de novo protein design", source="biorxiv", size=5)
reader.search("multimodal Alzheimer diagnosis", source="medrxiv", size=5)

# Opt-in to fine reranking
reader.search("transformer model", use_fine_rerank=True, size=10)

2. Trending and Popularity

Research is not only about what exists, but what is worth reading now.

deepxiv trending --days 7 --limit 30
deepxiv paper 2409.05591 --popularity

• trending finds the hottest recent papers from social signals
• --popularity gives paper-level propagation metrics such as views, tweets, likes, and replies

3. Web Search

deepxiv wsearch "karpathy"
deepxiv wsearch "karpathy" --json

Notes:

• deepxiv wsearch calls the DeepXiv web search endpoint
• each wsearch request costs 20 scores, other requests cost 1 score
• an auto-registered anonymous token gets 1,000 scores per day (~50 web searches/day)
• a token registered at data.rag.ac.cn/register gets 10,000 scores per day (~500 web searches/day)

4. Semantic Scholar Metadata by ID

deepxiv sc 258001
deepxiv sc 258001 --json

deepxiv sc fetches metadata using a Semantic Scholar paper ID.

Notes:

• this is useful when your workflow already has Semantic Scholar IDs
• DeepXiv will soon provide a Semantic Scholar search service that returns Semantic Scholar IDs directly

5. Biomedical Papers

deepxiv pmc PMC544940 --head
deepxiv pmc PMC544940

6. bioRxiv & medRxiv Preprints

✅ Available on PyPI (since deepxiv-sdk>=0.2.4). Just pip install -U deepxiv-sdk. Preprint search now shares the unified retrieve endpoint with arXiv.

# Search (see "Paper Search and Reading" above for more patterns)
deepxiv search "protein design" --biorxiv --limit 5
deepxiv search "Alzheimer" --medrxiv --date-from 2024-01

# Fetch a paper by DOI
deepxiv biorxiv 10.1101/2021.02.26.433129
deepxiv biorxiv 10.1101/2021.02.26.433129 --format text
deepxiv biorxiv 10.1101/2021.02.26.433129 --section Introduction,Methods

deepxiv medrxiv 10.1101/2025.08.11.25333149
deepxiv medrxiv 10.1101/2025.08.11.25333149 --format text

# Also works via --biorxiv / --medrxiv flags on the paper command
deepxiv paper 10.1101/2021.02.26.433129 --biorxiv
deepxiv paper 10.1101/2021.02.26.433129 --biorxiv --section Introduction

Example Agent Workflows

Workflow 1: Review recent hot papers

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

Then an agent can:

1. run --brief for each paper
2. run --head for the most promising ones
3. read only key sections
4. produce a report without manually opening dozens of papers

This workflow has already been written as a reusable skill. See skills/deepxiv-trending-digest/SKILL.md.

Workflow 2: Enter a new research topic

deepxiv search "agentic memory" --date-from 2026-03-01 --limit 100 --format json

Then an agent can:

1. batch-brief the results
2. prioritize papers with GitHub links
3. inspect experiments via --head
4. read Experiments / Results
5. turn datasets, metrics, and scores into a baseline table

This workflow is also available as a reusable skill. See skills/deepxiv-baseline-table/SKILL.md.

Built-in Deep Research Agent

If you do not want to compose the workflow manually, the CLI already includes a built-in research agent.

pip install "deepxiv-sdk[all]"
deepxiv agent config
deepxiv agent query "What are the latest papers about agent memory?" --verbose

If you already have your own agent stack, you can also just plug in the DeepXiv CLI skill and keep your own orchestration.

Agent Integration

DeepXiv is designed to work well inside Codex, Claude Code, OpenClaw, and similar agent runtimes.

MCP Server

Add to Claude Desktop MCP config file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "deepxiv": {
      "command": "deepxiv",
      "args": ["serve"],
      "env": {
        "DEEPXIV_TOKEN": "your_token_here"
      }
    }
  }
}

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.

Python Usage

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"])

# Switch sources without a separate method.
bio_hits = reader.search("de novo protein design", source="biorxiv", size=5)
med_hits = reader.search("multimodal Alzheimer diagnosis", source="medrxiv", size=5)

brief = reader.brief("2409.05591")
head = reader.head("2409.05591")
intro = reader.section("2409.05591", "Introduction")

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

Upgrade note (2026-04): the unified retrieve endpoint replaced the old Elasticsearch-style interface. See Search API changes below.

Roadmap

DeepXiv is moving toward an academic paper data interface at 100M+ scale.

The roadmap is:

1. Full arXiv coverage with T+1 automatic updates
2. anyXiv coverage, including bioRxiv, medRxiv, and similar repositories
3. Full open-access literature coverage

The metadata backbone will increasingly rely on Semantic Scholar metadata as the base layer, while continuously expanding coverage and enrichment quality.

Current Coverage

• ✅ arXiv - current primary source
• ✅ PubMed Central (PMC) - biomedical and life sciences
• ✅ bioRxiv / medRxiv - biology & medicine preprints (available in the PyPI release)
• 🔄 Semantic Scholar metadata integration - expanding as the metadata foundation

DeepXiv focuses on open-access literature so agents can work on unrestricted paper data instead of getting blocked by subscription barriers.

Complete API Reference

Search and Query

reader.search(
    query,
    size=10,                  # mapped to upstream top_k (1~100)
    offset=0,                 # 0~10000
    source="arxiv",           # "arxiv" | "biorxiv" | "medrxiv"
    categories=None,          # list[str]; filter only
    authors=None,             # list[str]; also influences ranking
    orgs=None,                # list[str]; also influences ranking
    min_citation=None,
    date_from=None,           # convenience (mapped to date_search_type)
    date_to=None,             # convenience (mapped to date_search_type)
    date_search_type=None,    # advanced: "between"|"exact"|"after"|"before"
    date_str=None,            # advanced: str or [start, end]
    use_fine_rerank=False,    # SDK default: off
)
reader.websearch(query)            # Web search (20 limit per request)
reader.semantic_scholar(sc_id)     # Metadata lookup by Semantic Scholar ID
reader.head(arxiv_id)              # Paper metadata and sections overview
reader.brief(arxiv_id)             # Quick summary (title, TLDR, keywords, citations, GitHub URL)
reader.section(arxiv_id, section)  # Read specific section
reader.raw(arxiv_id)               # Full paper
reader.preview(arxiv_id)           # Paper preview (~10k characters)
reader.json(arxiv_id)              # Complete structured JSON

The search response shape now matches the upstream retrieve endpoint:

{
  "status": "success",
  "total_count": 3,
  "result": [
    {
      "arxiv_id": "2506.18871",    // biorxiv_id / medrxiv_id when source != arxiv
      "title": "...",
      "score": 0.9475,
      "abstract": "...",
      "tldr": "...",
      "authors": [{ "name": "...", "orgs": ["..."] }],
      "url": "...",
      "date": "2025-06-23T17:38:54Z",
      "citation_count": 217,
      "categories": ["cs.CV"]
    }
  ]
}

Search API changes (2026-04)

The search backend moved from Elasticsearch to the unified /arxiv/?type=retrieve retrieval service. The SDK keeps parameter names where possible, but a few behaviors changed — please review before upgrading:

Parameter	Status	Notes
size	kept	Still works. Internally mapped to upstream top_k. You can also pass top_k= directly.
offset	kept	Now capped at 0~10000.
categories, authors, min_citation	kept	Same semantics.
source	new	"arxiv" (default), "biorxiv", "medrxiv". Replaces the separate bioRxiv / medRxiv search endpoints. reader.biomed_search() is now a thin wrapper around search(source=...).
orgs	new	Organization filter; also influences ranking.
date_search_type / date_str	new	Advanced date filter. Supports between / exact / after / before.
date_from / date_to	kept (mapped)	Automatically converted to date_search_type + date_str: both → between; only date_from → after; only date_to → before. Now also accepts YYYY / YYYY-MM (previously only YYYY-MM-DD).
use_fine_rerank	new	Upstream default is True; the SDK defaults to False so regular calls stay cheap. Set to True when you want better ordering.
search_mode / bm25_weight / vector_weight	deprecated	Accepted for backward compatibility but ignored (a warning is logged). Remove them from your code.
search_funcs	not exposed	The SDK always uses the full default index set (metadata + section + roc).
return_contents / return_roc	not exposed	Always disabled. Use reader.raw() / reader.section() / reader.json() to fetch content.

Response-shape migration:

• {total, took, results} → {status, total_count, result}
• Per-item ID field depends on source: arxiv_id / biorxiv_id / medrxiv_id
• Old paper["citation"] → new paper["citation_count"]

CLI equivalents:

• --limit → maps to size/top_k
• --offset, --authors, --orgs, --categories, --min-citations — direct filters
• --date-from / --date-to — legacy convenience, auto-mapped
• --date-search-type / --date-str — advanced (use --date-str twice for between)
• --use-fine-rerank — opt-in flag
• --biorxiv / --medrxiv — switch source
• --mode — deprecated, no-op

PMC (Biomedical Papers)

reader.pmc_head(pmc_id)            # PMC paper metadata
reader.pmc_full(pmc_id)            # Complete PMC paper JSON

bioRxiv / medRxiv

Preprint search now shares the unified retrieve endpoint with arXiv. Prefer reader.search(..., source="biorxiv" | "medrxiv"); biomed_search remains as a thin compatibility wrapper.

# Recommended: unified search
reader.search("protein design", source="biorxiv", size=10)
reader.search("Alzheimer", source="medrxiv", size=10)

# Per-paper data access (DOI-based) is unchanged
reader.biomed_data(source_id, source="biorxiv")
reader.biomed_data(source_id, source="biorxiv", data_type="section", section_names=["Introduction"])
reader.biomed_data(source_id, source="medrxiv", data_type="roc", roc_num=5)

# Legacy (still supported)
reader.biomed_search(query, source="biorxiv", top_k=10)

Agent (Optional)

from deepxiv_sdk import Agent

agent = Agent(api_key="your_openai_key", model="gpt-4")
answer = agent.query("What are the latest papers about agent memory?")
print(answer)

Token Management

deepxiv supports 4 ways to configure tokens:

1. Auto-registration (Recommended) - Automatically creates and saves on first use (1,000 requests/day)

deepxiv search "agent"

2. Using config command

deepxiv config --token YOUR_TOKEN

3. Environment variable

export DEEPXIV_TOKEN="your_token"

4. Command-line option

deepxiv paper 2409.05591 --token YOUR_TOKEN

Daily limits by token type:

Token type	Daily limit	How to get
Auto-registered (anonymous)	1,000 requests	Happens automatically on first CLI use
Registered token	10,000 requests	Sign up at data.rag.ac.cn/register
Custom / higher limit	Contact us	Email tommy[at]chien.io and describe your use case

Free Test Papers

These papers can be accessed without a token:

arXiv: 2409.05591, 2504.21776 PMC: PMC544940, PMC514704

MCP Tools

Available tools when using MCP Server:

Tool	Description
search_papers	Search arXiv papers
get_paper_brief	Quick summary
get_paper_metadata	Full metadata
get_paper_section	Read specific section
get_full_paper	Complete paper
get_paper_preview	Paper preview
get_pmc_metadata	PMC paper metadata
get_pmc_full	Complete PMC paper

Agent Usage (Optional)

The built-in ReAct agent can automatically search papers, read content, and perform multi-turn reasoning:

from deepxiv_sdk import Agent

agent = Agent(
    api_key="your_deepseek_key",
    base_url="https://api.deepseek.com/v1",
    model="deepseek-chat"
)

answer = agent.query("Compare key ideas in transformers and attention mechanisms")
print(answer)

Or via CLI:

deepxiv agent config  # Configure LLM API
deepxiv agent query "What are the latest papers about agent memory?" --verbose

Error Handling

deepxiv provides specific exception types:

from deepxiv_sdk import (
    Reader,
    AuthenticationError,  # 401 - Invalid or expired token
    RateLimitError,       # 429 - Daily limit reached
    NotFoundError,        # 404 - Paper not found
    ServerError,          # 5xx - Server error
    APIError              # Other API errors
)

try:
    paper = reader.brief("2409.05591")
except AuthenticationError:
    print("Please update your token")
except RateLimitError:
    print("Daily limit reached")
except NotFoundError:
    print("Paper not found")
except APIError as e:
    print(f"API error: {e}")

Troubleshooting

Q: Do I need a token to use? A: No. Some papers are free to access. Search and some content require a token, but it's auto-created on first use.

Q: What's the maximum search results? A: 100 per request. Use offset parameter for pagination.

Q: How to handle timeouts? A: Reader automatically retries (max 3 times) with exponential backoff. You can customize:

reader = Reader(timeout=120, max_retries=5)

Q: Can I cache paper content? A: Yes. After getting content with reader, cache locally to database or file system.

Q: Which LLMs does the agent support? A: Any OpenAI-compatible API (OpenAI, DeepSeek, OpenRouter, local Ollama, etc.).

Examples

See examples/ directory:

• quickstart.py - 5-minute quick start
• example_reader.py - Basic Reader usage
• example_agent.py - Agent usage
• example_advanced.py - Advanced patterns
• example_error_handling.py - Error handling examples

License

MIT License - see LICENSE file

Support

• 🐛 GitHub Issues: https://github.com/qhjqhj00/deepxiv_sdk/issues
• 📚 API Documentation: https://data.rag.ac.cn/api/docs
• 📧 Higher Limits: Register at data.rag.ac.cn/register for 10,000 requests/day, or email tommy[at]chien.io to describe your use case for a custom limit