Deepxiv CLI

DeepXiv CLI Skill

🚀 30-Second Overview

What: Access open access academic papers (arXiv, PMC, bioRxiv, medRxiv) through a powerful CLI tool.

What you can do:

• 💬 Ask a question and get a cited answer, from arXiv or the web (deepxiv ask)
• 🔍 Search papers with semantic search
• 📄 Read papers section-by-section (save tokens!)
• 🧠 Use built-in AI agent for analysis
• 📊 Access biomedical literature (PMC)

When to use: User wants to search, read, or analyze academic papers

🎯 Command Selection Guide

I want to... → Use this command

📚 Core Commands Reference

0. Ask a Question (deepxiv ask)

When to use: The goal is an answer, not a specific paper's text. This runs the whole search → judge → read loop server-side and returns prose with real citations. Reach for it before hand-rolling search + paper --section.

# arXiv (default) — methods, numbers, experimental results
deepxiv ask "what speedup does speculative decoding report on HumanEval"
deepxiv ask "对比几种缓解 MoE 路由崩塌的方法" --effort high

# --web — current events, products, pricing, anything non-academic
deepxiv ask "Anthropic Claude API pricing tiers" --web
deepxiv ask "NeurIPS 2025 best paper winner" --web --search-type news
deepxiv ask "retrieval evaluation methodology" --web --search-type scholar

Requires a registered key. The auto-registered SDK token returns 403 — users must register at https://data.rag.ac.cn/register. Every account gets 300 agentic calls/day free, from a quota pool separate from the general daily limit (so a 429 here does not block search or paper). If a call 403s, tell the user to register rather than retrying.

Takes 8–30s end to end; the answer streams, so text appears well before that.

Writing the query — this matters more than any flag:

• Be specific. "what compression ratio does KV cache eviction report on LongBench" works; "kv cache" does not.
• Ask for numbers explicitly ("what speedup", "which benchmark") to make it read paper bodies rather than abstracts.
• Put scope in the query text: year, venue, category, author, min citations.
• Chinese queries work directly; the answer comes back in the query's language.
• If the answer misses, rephrase — don't just raise --effort. Effort adds reading rounds but cannot redirect first-round recall.

Reading the output:

• Answer → stdout; sources and progress → stderr. deepxiv ask "..." > answer.md captures only the answer.
• Every [arXiv:ID] is real — the service is instructed never to invent one, and says "no relevant papers" instead of fabricating. Keep those IDs in what you report back to the user.
• The sources list is the retrieval set, a superset of what's cited. The CLI shows only cited sources; --all-sources shows the rest.
• On --web, weigh evidence strength. The service reads only cached page bodies and never fetches live, so uncached pages contribute just a search snippet. The CLI marks pages read in full (📄) versus snippet-only (🔗), and the answer flags snippet-only claims. Pass that caveat on to the user instead of presenting everything as equally solid.
• On a truncation warning, do not treat the answer as complete — re-run with a larger --max-answer-tokens or a narrower query.
• Follow up on any cited paper with the normal commands: deepxiv paper <id> --section Method.

1. Search Papers (deepxiv search)

When to use: Finding relevant papers on a topic

# Basic search (quick)
deepxiv search "agent memory" --limit 5

# Advanced search with filters
deepxiv search "transformer" --format json
deepxiv search "LLM" --categories cs.AI,cs.CL --min-citations 100

# Date range
deepxiv search "vision models" --date-from 2024-01-01 --date-to 2024-12-31

Expected output: List of papers with title, arxiv_id, score, citation count Token cost: ≈500-1000 tokens per search Tips:

• Use --limit 3-5 for quick overview
• Use --format json for downstream processing
• Add --use-fine-rerank when ranking quality matters more than latency

2. Read arXiv Papers (deepxiv paper)

When to use: Understanding specific papers (choose format by depth needed)

Option A: Quick Summary (30 seconds)

deepxiv paper 2409.05591 --brief
# ✓ Best for: Quick understanding
# → Title, TLDR, keywords, citation count
# Token cost: ~500 tokens
# Use when: "What's this paper about?"

Option B: Paper Structure (2 minutes)

deepxiv paper 2409.05591 --head
# ✓ Best for: Understanding paper organization
# → Metadata, all sections with summaries
# Token cost: ~1-2k tokens
# Use when: "What does this paper contain?"

Option C: Quick Scan (3 minutes)

deepxiv paper 2409.05591 --preview
# ✓ Best for: Fast preview without full read
# → First ~10k characters
# Token cost: ~2k tokens
# Use when: "Let me scan the introduction"

Option D: Specific Section (1 minute)

deepxiv paper 2409.05591 --section Introduction
# ✓ Best for: Reading one section in detail
# → Full section content
# Token cost: ~1-5k tokens (varies by section)
# Use when: "I only need the introduction"

Option E: Complete Paper (for deep analysis)

deepxiv paper 2409.05591
# or explicitly: deepxiv paper 2409.05591 --raw
# ✓ Best for: Full text analysis
# → Complete paper in markdown
# Token cost: ~10-50k tokens
# Use when: "I need to thoroughly analyze this"

3. Access Biomedical Literature (deepxiv pmc)

When to use: Need medical, biology, or biomedical research

# Quick metadata
deepxiv pmc PMC544940 --head

# Full paper
deepxiv pmc PMC544940

Expected output: Paper metadata or complete paper JSON Token cost: ≈1-2k tokens (metadata), ≈10-30k tokens (full)

4. Manage Token (deepxiv token)

When to use: Check your API token status

deepxiv token
# Shows current token and daily limit info

Note: Token auto-registers on first use, saves to ~/.env