FengYM666/AI-Research-OS

Paper Search Specification (Paper Search)

The user's primary job is academic research. When searching for papers for them, accuracy > quantity > speed.

Iron Rule: No Fabrication

1. Every recommended paper must come from an actual record retrieved in the current session (API response or webpage content). Do not list papers from training memory—names, titles, and years from memory are extremely prone to being garbled.
2. The five essential citation elements must all be complete and come from the same actual record: title, authors, year, publication venue, accessible link.
3. If nothing is found, simply say "not found"; if relying on memory without verification, mark it as [Unverified] and explain why.
4. Provide the exact search terms used so the user can reproduce and dig deeper.
5. Distinguish between the formally published version and the arXiv preprint; if titles match, prioritize the formally published information.

Primary Tool: paper-search MCP (Configured)

ZCode is already connected to the paper-search MCP server (openags/paper-search-mcp, 57 tools). Prefer calling its tools for retrieval instead of using curl:

• search_papers (cross-database unified search), search_arxiv, search_openalex, search_dblp, search_semantic, search_google_scholar, search_crossref, etc. — for searches across 20 sources
• download_arxiv / read_arxiv_paper, etc. — for downloading PDFs, extracting text (each source has corresponding tools)
• Google Scholar search is already configured with a proxy (PAPER_SEARCH_MCP_GOOGLE_SCHOLAR_PROXY_URL); if the proxy is not active, GS will fail—just switch to another source

When MCP is unavailable (e.g., the server is not connected), fall back to the manual curl approach below.

Sources and Connectivity (Tested Locally on 2026-09)

Directly accessible (no proxy needed from terminal):

Source	Purpose	Entry Point
OpenAlex API	Primary choice: free, no key required, covers 250M+ papers, filterable by date/venue/citations	https://api.openalex.org/works?search=<term>&per-page=10, add &filter=from_publication_date:YYYY-01-01 to limit by year
arXiv API	Preprints, where CV/AI new papers are first published	https://export.arxiv.org/api/query?search_query=all:<term>&max_results=10&sortBy=submittedDate
CVF Open Access	Official full text for CVPR / ICCV / ECCV, including workshops	openaccess.thecvf.com
dblp API	The gold standard for verifying the actual publication venue and year of a paper	https://dblp.org/search/publ/api?q=<title>&format=json
Papers with Code	Papers + official code + SOTA leaderboards	paperswithcode.com
Domain official libraries	ACL Anthology (NLP), proceedings.neurips.cc, proceedings.mlr.press (ICML), OpenReview (ICLR)	respective official websites

Requires a proxy (first run export https_proxy=http://127.0.0.1:7892 in terminal; port may vary based on current registry):

Source	Purpose	Note
Google Scholar	Widest coverage, citation counts, citation tracking, "related articles" expansion	Priority: use control-browser skill to open a real browser and search (system proxy takes effect automatically, can bypass CAPTCHAs); curl scraping easily triggers CAPTCHA
Semantic Scholar API	Citation network analysis, TLDR summaries	No key → heavy rate limiting (easily hits 429); usable for small requests only — use as fallback

Workflow

1. Deconstruct search terms: Break the user's request into 2–4 English search queries (academic databases mostly only recognize English; translate Chinese keywords first).
2. Initial filter: Search OpenAlex + arXiv, select candidates based on relevance, year, and citation count.
3. Verify venue: For papers claimed to be published at CVPR/ICCV, etc., check against dblp or the conference's official website — this is the specific step to kill "plausible-looking fake papers."
4. Find code/SOTA: Papers with Code.
5. Expand with classics: Google Scholar sorted by citations + "related articles" + crawl reference lists.
6. Before output, check each entry against the five elements; if any is missing, go back and look it up.

Output Format

Title	Authors	Conference/Journal + Year	Link	Relevance (one sentence)

Following the table: search terms used this session; if there is official code, include the code repository link; indicate which entries were not yet verified.