Retrieval Sources Reference / 检索来源参考

Use this file in Module 2 to:

Detect the query domain
Activate only the matching source packs
Look up venue lists and scoring rules

Domain Router / 领域路由（必读）

After Module 1 query rewrite, detect 1–2 primary domains, then load the corresponding source packs below.

Domain ID	Trigger keywords (CN / EN examples)	Activate packs
`ai_cs`	大模型, LLM, CV, NLP, 强化学习, transformer, deep learning, computer vision, IR, data mining	A + B (+ A-CN when CN/国产/中文大模型)
`math_physics`	数学, 物理, 粒子物理, nonlinear, quantitative biology	A(arXiv) + C + E(CERN/OSTI if energy/aero)
`chemistry`	化学, 分子, CAS, 化工, chemistry, reagent	D + A(arXiv q-bio/cond-mat if needed)
`materials`	材料, 纳米, materials science, NIMS	D + E(OSTI) + A
`biomed`	医学, 生物, PubMed, clinical, genome, drug	F + A(q-bio) + H(SciELO/DOAJ)
`education`	教育, pedagogy, curriculum, edtech	F(ERIC) + H + I
`humanities_ss`	哲学, 历史, 政治, 社会, 法学, 新闻, 人文社科	I + H + J(newspapers if media history)
`economics`	经济, 金融, 管理, marketing, SWOT, econ	I(EconLit/SSRN-class) + K(EBSCO-class)
`engineering_energy`	能源, 航空航天, 工程, NASA, DOE	E + A + H
`patent_ip`	专利, patent, IP	L only (+ domain pack if tech context)
`thesis`	学位论文, thesis, dissertation, 硕博	M + domain pack
`general_oa`	开放获取, OA, open access, 找不到领域	H + A + N
`cross_domain`	跨学科 / unclear	A + H + N → then refine

Routing rules / 路由规则

Always run Pack A for STEM queries that may have preprints (AI/CS/math/physics/bio).
Always keep English search terms for international packs, even in /query_chinese mode.
Cap active packs at 3 unless the user asks for exhaustive coverage.
If domain confidence is low → use general_oa + ask one clarifying question.
Do not activate shadow libraries / pirate ebook sites for skill retrieval.
CN LLM labs (A-CN): if query mentions 国产/中文大模型/Qwen/DeepSeek/GLM/InternLM/混元/豆包/文心/星火 etc., activate A-CN Tier-1 matching orgs (2–4); if query is a broad CN-LLM survey, also scan A-CN Discovery.
Named-model force map: see A-CN force triggers.

Priority scholarly APIs (2.14 routing)

When Pack A / F / H is active (or general_oa / cross_domain), prefer these hubs (web UI and/or API; compose article-mcp if configured):

Source	Use	How
Semantic Scholar	AI/CS citation + metadata	https://www.semanticscholar.org/ · optional `S2_API_KEY`
OpenAlex	Broad scholarly graph / open metadata · default API lane	https://openalex.org/ · API below · polite pool
Crossref	DOI resolve → BibTeX enrich	https://api.crossref.org/
PubMed / NCBI	Biomed (`biomed` domain → Pack F)	https://pubmed.ncbi.nlm.nih.gov/
OpenReview	ICLR & drafts	https://openreview.net/
DBLP	CS venue metadata	https://dblp.org/

Do not treat source count as a KPI; activate ≤3 packs unless user asks exhaustive.

OpenAlex API recipes / OpenAlex 检索配方（必用）

Base URL: https://api.openalex.org
Docs: https://docs.openalex.org/ · UI: https://openalex.org/
Polite pool (required for reliable rate limits): append mailto=your@email.com on every request, or set User-Agent to a contactable string with email.

1) Basic search / 基础检索

curl "https://api.openalex.org/works?search=math+tutoring&per-page=20&mailto=you@example.com"

search — free-text (title/abstract/fulltext index); URL-encode spaces as + or %20
per-page — page size (default 25, max 200); Module 2 usually uses 20–50 then score
Cursor pagination: cursor=* then follow meta.next_cursor

2) Filtered + sorted search / 带筛选的检索

curl "https://api.openalex.org/works?search=math+tutoring&filter=publication_year:2020-2026&sort=relevance_score:desc&mailto=you@example.com"

Useful filter fragments (combine with ,):

Filter	Example	When
Year range	`publication_year:2020-2026`	Recency / 「最新」
Single year	`publication_year:2024`	Snapshot year
OA only	`is_oa:true`	Need free full text
Type	`type:article` / `type:preprint`	Drop misc
Concept	`concepts.id:C41008148`	Topic focus (lookup concept first)
Author / institution	`authorships.author.id:A…` / `authorships.institutions.id:I…`	Expert / lab dig

Useful sort: relevance_score:desc (default for search) · publication_date:desc · cited_by_count:desc

3) Work detail / 单篇详情

curl "https://api.openalex.org/works/W2206575885?mailto=you@example.com"

Also accepts DOI / OpenAlex URL forms, e.g. /works/https://doi.org/10.7717/peerj.4375.

Field → report mapping / 字段映射

OpenAlex field	Report / score use
`id` / `ids.openalex`	Dedup key `OpenAlex:W…`
`display_name`	Title
`publication_year` / `publication_date`	Recency
`doi` / `ids.doi`	Dedup + Crossref enrich
`ids.arxiv` / `ids.pmid`	Cross-link Pack A / F
`cited_by_count`	Importance signal (secondary)
`primary_location` / `best_oa_location`	Venue + OA PDF URL
`authorships`	Authors / affiliations
`abstract_inverted_index`	Reconstruct abstract for Similarity
`concepts`	Domain tags
`type`	article / preprint / …

Agent usage rules / 使用规则

For Pack A / H / general_oa / cross_domain, run at least one OpenAlex works?search= call with the English Keyword query (plus year filter when latest-intent).
Multi-strategy (paper-search-pro): for high-stakes / deep queries, fan out 3 OpenAlex sorts on the same cleaned topic — relevance_score:desc, publication_date:desc, cited_by_count:desc — then RRF-fuse (do not treat them as unrelated topics).
Prefer OpenAlex for broad scholarly recall and DOI/OA links; prefer arXiv / HF / PwC for ML preprint + code; prefer S2 when citation neighborhood is needed.
Deduplicate: DOI (normalized) → arXiv ID (strip vN) → PMID → OpenAlex ID → title+year (see Scoring Reference / wiki_bridge.rrf).
Do not invent OpenAlex IDs; only cite works returned by the API or UI.
If API 403/429: retry with mailto=, reduce per-page, or fall back to https://openalex.org/ web search.

Pack A — AI / Computer Science Core（计算机 · AI 主库）

Default pack for Paper_Rec. Highest priority for ML/AI/CV/NLP/systems queries.

Source	URL	Search tip
OpenAlex	https://openalex.org/ · API `https://api.openalex.org/works`	Broad graph; see OpenAlex API recipes
arXiv	https://arxiv.org/search/	Filter `cs.LG` `cs.CV` `cs.CL` `cs.AI` `cs.IR` `stat.ML`
Hugging Face Papers	https://huggingface.co/papers	Daily trending ML papers
Papers With Code	https://paperswithcode.com/	SOTA + code links
GitHub	https://github.com/search	Implementations, stars, topics
Semantic Scholar	https://www.semanticscholar.org/	Citation graph · API-friendly metadata
DBLP	https://dblp.org/	CS bibliography / venue lookup (also https://dblp.uni-trier.de/)
OpenReview	https://openreview.net/	ICLR & peer-review drafts
Connected Papers	https://www.connectedpapers.com/	Graph neighborhood from a seed paper
AMiner	https://www.aminer.cn/	CN-friendly scholar/paper graph (alt https://mrt.aminer.cn/)
Google Scholar	https://scholar.google.com/	Broad citation search; prefer browser (anti-bot)
ACL Anthology	https://aclanthology.org/	NLP proceedings
CVF Open Access	https://openaccess.thecvf.com/	CVPR / ICCV / ECCV

CCF-A venues (AI/CS)

Abbrev	Field
NeurIPS, ICML, ICLR	ML
CVPR, ICCV	Vision
ACL, AAAI, IJCAI	NLP / AI
WWW, SIGKDD, SIGIR	Web / DM / IR

CCF-B (selective)

EMNLP, NAACL, ECCV, BMVC, COLING, WSDM, CIKM

Conference / journal hubs（开放全文）

NeurIPS: https://papers.nips.cc/ · https://proceedings.neurips.cc/
ICML: https://proceedings.mlr.press/
ICLR: https://openreview.net/group?id=ICLR.cc
JMLR: https://jmlr.csail.mit.edu/papers/ · https://www.jmlr.org/

Venue planning meta（非正文检索；投稿/档期时启用）

Source	URL	Use
CCF AI 评估列表	https://www.ccf.org.cn/Academic_Evaluation/AI/	CCF 分区参考
CCF Deadlines	https://ccfddl.github.io/	截稿日历
AI Deadlines	https://aideadlin.es/?sub=ML,CV,NLP	ML/CV/NLP 截稿
会伴 myhuiban	https://www.myhuiban.com/?lang=zh_cn	国内会议日历

Pack A-CN — Chinese LLM Labs（国产大模型 · 论文发现补充）

Source catalog distilled from awesome-LLMs-In-China.
Purpose: catch latest CN foundation-model tech reports / arXiv / GitHub releases that global indexes may lag on.

A-CN Discovery（发现层）

Source	URL	When to use
awesome-LLMs-In-China (site)	https://free-gpt.github.io/awesome-LLMs-In-China/	Broad 国产/中文大模型 survey; discover new teams
awesome-LLMs-In-China (repo)	https://github.com/wgwang/awesome-LLMs-In-China	Track updates / Issues
awesome-open-foundation-models	https://github.com/wgwang/awesome-open-foundation-models	Open foundation model index
awesome-LLM-benchmarks	https://github.com/wgwang/awesome-LLM-benchmarks	Benchmarks → reverse-find papers
ModelScope 魔搭	https://www.modelscope.cn/	CN model cards often link tech reports
BAAI Model Hub	https://model.baai.ac.cn/	Aquila / FlagAI ecosystem docs

Default: enable Discovery only for queries about 国产大模型 / 中文 LLM 综述 / CN foundation models — not every ai_cs query.

A-CN Tier-1 Labs（默认高优先）

For matching keywords, search GitHub org + HF org + official tech report/blog; always prefer arXiv/PDF primary link when found.

Team	Models / aliases	Primary channels
DeepSeek 深度求索	DeepSeek, R1, Coder, V	https://github.com/deepseek-ai · https://huggingface.co/deepseek-ai
Alibaba Qwen / 通义 / 达摩	Qwen, Tongyi, Qwen-VL/Audio	https://github.com/QwenLM · https://huggingface.co/Qwen · https://tongyi.aliyun.com/
Zhipu / THUDM 智谱	ChatGLM, GLM, 清言	https://github.com/THUDM · https://huggingface.co/THUDM · https://chatglm.cn/
Shanghai AI Lab 书生	InternLM, InternVL	https://github.com/InternLM · https://intern-ai.org.cn/
BAAI 智源	Aquila, Emu, FlagAI, 悟道	https://github.com/FlagAI-Open · https://huggingface.co/BAAI · https://model.baai.ac.cn/
OpenBMB / 面壁	CPM, MiniCPM	https://github.com/OpenBMB · https://huggingface.co/openbmb · https://modelbest.cn/
01.AI 零一万物	Yi	https://github.com/01-ai/Yi · https://huggingface.co/01-ai
Moonshot 月之暗面	Kimi, Moonshot	https://www.moonshot.cn/ · official blog / tech reports
Baichuan 百川	Baichuan	https://github.com/baichuan-inc · https://huggingface.co/baichuan-inc
Tencent Hunyuan 混元	Hunyuan	https://hunyuan.tencent.com/ · Tencent AI Lab releases
ByteDance / 豆包 / Seed	Doubao, Seed, 云雀	https://www.doubao.com/ · ByteDance / Seed research posts
IDEA CCNL	姜子牙, Fengshenbang, Ziya	https://fengshenbang-lm.com/ · https://huggingface.co/IDEA-CCNL

A-CN Tier-2 Labs（按需）

Team	Models	Channels	Activate when
iFlytek 讯飞	星火 Spark	https://xinghuo.xfyun.cn/	星火 / 讯飞 / speech+LLM
Huawei Noah / 盘古 / 鹏城	PanGu	https://www.huaweicloud.com/product/pangu.html · openi.pcl.ac.cn	盘古 / 华为 / 工业大模型
SenseTime 商汤	商量, SenseNova	https://chat.sensetime.com/ · SenseTime research	商汤 / 多模态工业
Baidu ERNIE 文心	文心一言, ERNIE	https://yiyan.baidu.com/ · Baidu Research	文心 / ERNIE / 百度
Fudan OpenLMLab / DISC	MOSS, FinLLM	https://github.com/OpenLMLab	MOSS / 复旦 LLM
HKUST LMFlow	Robin, LMFlow	GitHub LMFlow / related papers	LMFlow / 训练框架
XVERSE 元象	XVERSE	https://github.com/xverse-ai	XVERSE
Kunlun Skywork 天工	Skywork	HF/GitHub Skywork	Skywork / 天工
vivo BlueLM	BlueLM	HF/GitHub BlueLM	BlueLM
Inspur 源	源 2.0	Inspur open releases	浪潮 / 源大模型

A-CN Academic labs（子领域）

Lab	Focus	Note
THUDM (Tsinghua)	GLM lineage	Overlaps Zhipu; use arXiv + GitHub
PKU groups	ChatLaw, CodeShell, Chat-UniVi	law / code / multimodal
ZJU NLP / OpenKG	KnowLM, 智海系列	knowledge / legal / edu
SJTU BAI / AI Institute	白玉兰, K2, GeoLLaMA	science LLM
HIT	本草, 活字	medical / Chinese LLM
CASIA	紫东·太初	full-modality
HUST / SYSU	Monkey, Firefly	multimodal SFT series

Vertical medical/legal/finance LLMs: prefer parent domain packs (biomed / humanities_ss / economics) plus the named lab channel above.

A-CN force triggers / 强制触发

Query contains	Force-open channels
DeepSeek, R1, deepseek	DeepSeek GitHub/HF + arXiv
Qwen, 通义, Tongyi	QwenLM + HF Qwen (prefer Qwen3.7 > 3.6 > 3.5 > 3 > 2.5; include official blog/docs if no TR yet)
GLM, ChatGLM, 智谱, 清言	THUDM + Zhipu
InternLM, 书生, InternVL	InternLM / Shanghai AI Lab
Aquila, 悟道, BAAI, FlagAI	BAAI hub + FlagAI
MiniCPM, CPM, 面壁, OpenBMB	OpenBMB
Yi- , 零一	01-ai/Yi
Kimi, Moonshot, 月之暗面	Moonshot official
Baichuan, 百川	baichuan-inc
混元, Hunyuan	Tencent Hunyuan
豆包, Seed, 云雀	ByteDance / Doubao
姜子牙, Fengshenbang, Ziya	IDEA-CCNL
国产大模型, 中文大模型综述	A-CN Discovery + Tier-1 sample (4+)

Rules: pick 2–4 Tier-1 orgs max per query (unless survey); always verify blog claims with arXiv/PDF; GitHub search may 429 — fall back to org page + HF model card + web search site:github.com/org.

Pack B — AI Industry & Trend（产业 · 趋势补充）

Use when Pack A hits <30 or user wants latest industrial work.

Global labs

Source	URL
Google Research	https://research.google/
Meta AI	https://ai.meta.com/research/
Microsoft Research	https://www.microsoft.com/en-us/research/
OpenAI Research	https://openai.com/research
DeepMind	https://deepmind.google/research/
Anthropic	https://www.anthropic.com/research
NVIDIA Research	https://research.nvidia.com/

CN industry portals（与 A-CN 互补；偏产品/发布，论文需回链）

Source	URL
阿里达摩院 / 通义	https://tongyi.aliyun.com/ · DAMO research posts
腾讯 AI Lab / 混元	https://hunyuan.tencent.com/
字节跳动 / 豆包	https://www.doubao.com/
华为诺亚 / 盘古	https://www.huaweicloud.com/product/pangu.html
百度文心	https://yiyan.baidu.com/
讯飞星火	https://xinghuo.xfyun.cn/
商汤	https://www.sensetime.com/

CN media (trend only → always verify with primary paper): 机器之心 · 量子位 · 阿里达摩院 · 腾讯AI Lab · 字节跳动技术 · 华为诺亚方舟

Influencer examples: CV → Kaiming He; RL → Levine / Abbeel; expand 2–3 leaders per subfield.

Pack C — Math / Physics / Preprint STEM（数理 · 预印本）

Source	URL	Focus
arXiv	https://arxiv.org/	`math.` `physics.` `hep-` `nlin.` `q-bio.*`
CERN Document Server	https://cds.cern.ch/	Particle physics (browser may work if API/script 403)
Science.gov	https://www.science.gov/	US federal STEM aggregation

Pack D — Chemistry / Materials（化学 · 材料）

Source	URL	Focus	Notes
ChemBlink	https://www.chemblink.com/	Chemical product / CAS lookup	Not a paper DB — use for substance metadata
Chemistry Central / Springer Chem	https://www.chemistrycentral.com/	Chem OA (may redirect Springer)	Prefer DOAJ + publisher OA
SciELO	https://www.scielo.org/	Regional OA journals	Good for LatAm chem/life sci
DOAJ	https://doaj.org/	OA journal/article search	Filter Chemistry / Materials
OSTI	https://www.osti.gov/	DOE reports, materials, energy	Full-text tech reports
NIMS (JP)	https://www.nims.go.jp/eng/	Materials research	Site may timeout; retry or use Google site:

Do not treat ChemBlink as a literature engine; pair it with DOAJ/OSTI/SciELO for papers.

Pack E — Engineering / Energy / Aerospace（工程 · 能源 · 航天）

Source	URL	Focus
NASA NTRS	https://ntrs.nasa.gov/	Aerospace reports & conference papers
OSTI / DOE	https://www.osti.gov/	Energy, materials, engineering reports
Science.gov	https://www.science.gov/	Multi-agency STEM
NSTL	https://www.nstl.gov.cn/	CN sci-tech literature hub

Pack F — Biomed / Life Sciences / Education（生医 · 教育）

Source	URL	Focus	Access
ERIC	https://eric.ed.gov/	Education literature	Free abstracts; some full text
Local PubMed portal	https://www.yuntsg.com/	PubMed-style biomedical search	Registration / trial
SciELO	https://www.scielo.org/	Life/health OA	Free
DOAJ	https://doaj.org/	OA medical journals	Free
Bioline	https://www.bioline.org.br/	Developing-country OA journals	Free
HighWire (Stanford)	https://highwire.stanford.edu/	Biomed publisher portal	Often unstable / 502

Prefer ERIC for education; prefer PubMed-class + DOAJ/SciELO for biomed.

Pack G — reserved (AI labs already in Pack B)

(Merged into Pack B — kept ID for router compatibility.)

Pack H — Open Access Aggregators（开放获取聚合）

Activate for any domain when needing free full text or OA coverage.

Source	URL	Strength
OpenAlex	https://api.openalex.org/works	Scholarly graph + OA locations; filter `is_oa:true`
DOAJ	https://doaj.org/	Quality-controlled OA journals
OALib	https://www.oalib.com/	Broad OA article search (CN-friendly)
Socolar	https://www.socolar.com/	OA discovery + full-text links
SciELO	https://www.scielo.org/	Regional OA
Open J-Gate	http://www.openj-gate.com/	OA journal portal (quality uneven)
Cambridge Repository	https://www.repository.cam.ac.uk/	Institutional OA
NAP	https://www.nationalacademies.org/publications/all	Free academy reports/books

Pack I — Humanities / Social Sciences（人文社科）

Source	URL	Focus	Access
NCPSSD	https://www.ncpssd.org/	国家哲学社会科学文献中心	Free / open portal
CSSCI	http://cssci.nju.edu.cn/	Chinese social sciences citation	Limited / institution
CSSN	https://www.cssn.cn/	中国社会科学网（动态/资讯）	Free (not a paper DB)
CNKI	https://www.cnki.net/	中文期刊/硕博/会议	Paywall / campus
Wanfang	https://www.wanfangdata.com.cn/	中文综合学术库	Paywall / campus
RDFYBK	http://www.rdfybk.com/	人大复印报刊资料	Subscription
Xinhua Wenzhai DB	https://www.xinhuawz.com/	新华文摘	Subscription
NTU E-Journal	https://ejournal.press.ntu.edu.tw/	台大学术期刊	Open / campus
JSTOR	https://www.jstor.org/	Humanities & social sciences	Subscription
SSRN-class	https://www.ssrn.com/	Working papers (econ/law/ss)	Often 403 via script; browser OK

Pack J — Media / Newspaper Archives（报刊史料）

Only when query is about journalism history, media studies, or primary news sources — not for method papers.

Source	URL
人民日报图文库	https://data.people.com.cn/rmrb/
光明日报库	http://epaper.gmw.cn/gmrbdb/
大公报(1902–1949)	https://tk.cepiec.com.cn/tknewsc/tknewskm
中国近代报刊库	https://tk.cepiec.com.cn/SP/

Pack K — Business / Economics Databases（经管）

Source	URL	Notes
EconLit (AEA)	https://www.aeaweb.org/econlit/	Index; full search usually via campus
EBSCO	https://www.ebscohost.com/	Marketing page; real search needs institution
ScholarVox	https://www.scholarvox.com/	Business ebooks; institution
SSRN	https://www.ssrn.com/	Preprints in econ/finance/law

Pack L — Patents（专利）

Source	URL
SooPAT	https://www.soopat.com/Home/Index

Use alone or alongside ai_cs / engineering_energy when the query is IP-focused.

Pack M — Theses / Dissertations（学位论文）

Source	URL	Region
theses.fr	https://theses.fr/	France
DART-Europe	https://www.dart-europe.org/basic-search.php	EU theses portal (may block bots; browser OK)
CALIS OPAC	https://opac2.calis.edu.cn/	CN union catalog
UCDR S / 全国图书馆参考咨询联盟	http://www.ucdrs.superlib.net/	Document delivery
UMI / ProQuest-class	https://wwwlib.umi.com/	Intl theses (access limited)
CNKI / Wanfang degree DBs	cnki / wanfang	CN theses (paywall)

Pack N — CN National Sci-Tech & Libraries（国内科技文献枢纽）

Source	URL	Role
NSTL	https://www.nstl.gov.cn/	National sci-tech literature
CALIS	https://opac2.calis.edu.cn/	Higher-ed union catalog
Tsinghua Library portal	https://lib.tsinghua.edu.cn/	Gateway (not a paper DB itself)
paper.edu.cn	https://www.paper.edu.cn/	CN sci-tech paper online
Nature (journal hub)	https://www.nature.com/	Flagship STM; mostly paywalled

Explicitly Excluded / 明确排除（不要作为 skill 检索源）

Category	Examples	Reason
Shadow libraries	Library Genesis, Z-Library, BookSC/BookZZ, Sci-Hub (`sci-hub.*`)	Copyright risk; unstable
Scholar / PDF mirrors	`scholar.scqylaw.com`, `gfsoso.99lb.net`, `ac.scmor.com`	Unofficial; often broken / piracy-adjacent
Predatory OA catalogs	SCIRP journal listing as sole source	Quality uneven
Writing-only hubs	home-for-researchers as sole “search”	Not a paper corpus
Doc pirate / 文库破解	hi138, ebuymed, bingdian-style	Not scholarly; unreliable
Search bookmark hubs	虫部落 chongbuluo	Meta-directory, not a corpus
Broken / archive-only	CiteSeerX live index, many 404 portals	Not searchable anymore
Product-only chem shops	ChemBlink as sole source	Metadata ≠ papers

Pack A supplements (Connected Papers / AMiner / Scholar / JMLR / deadline hubs) curated from Awesome-Resources/Research after a 2026-07 reachability probe; pirate & mirror sites excluded.

Domain → Pack Quick Map / 速查表

ai_cs              → A, B  (+ A-CN if CN LLM keywords)
math_physics       → A, C
chemistry          → D, H
materials          → D, E, H
biomed             → F, H, A(q-bio)
education          → F(ERIC), H, I
humanities_ss      → I, H
economics          → K, I, H
engineering_energy → E, N, H
patent_ip          → L (+ parent domain pack)
thesis             → M (+ parent domain pack)
general_oa         → H, A, N
cross_domain       → A, H, N
cn_llm_survey      → A, A-CN(Discovery+Tier-1), B

Scoring Reference / 打分参考

Venue tier (Importance)

Tier	Examples	Boost
Tier-1 (AI/CS)	NeurIPS, ICML, ICLR, CVPR, ICCV, ACL, AAAI	+3
Tier-1 (field)	Field-top journals/conferences for non-CS domains	+3
Tier-2	EMNLP, ECCV, NAACL, strong OA journals	+2
Tier-3	arXiv preprint, workshop, tech report	+0–1
Industry flagship	Google / Meta / OpenAI flagship	+2
National hub hit	NSTL / NCPSSD / ERIC verified record	+1

Recency default

Default: prefer last 12 months for ranking; last 24 months still eligible.
If query has 最新 / latest / recent: apply latest-intent hard rules in \SKILL.md\ §2.3 — newer same-family versions must outrank older citation magnets.
Model-family order example (Qwen): �.7 > 3.6 > 3.5 > 3 > 2.5 > 2. Product release notes count when no arXiv TR exists.
Keep seminal older works only as labeled baselines, not as「最新」answers.

Deduplication keys

DOI (normalized; strip https://doi.org/, arXiv DOI vN) — primary wiki_bridge.rrf._doc_key
arXiv ID (strip vN)
PMID
OpenAlex work ID (W…)
Normalized title + year
Same PDF URL
Same model family + version (keep newest)

When merging across lanes/sources, keep all lane tags (_lanes / every source); never drop a keep-rule lane just because another copy scored higher.

When latest-intent is on, keep the newest version even if an older duplicate has a richer venue.

Module-2 checklist (domain-aware)

- [ ] Detect Domain ID(s) from rewritten query
- [ ] If CN LLM keywords → announce A-CN labs (2–4) or Discovery
- [ ] Announce: Domain = {id} → Packs = {list}
- [ ] Search Pack sources with Specific + Keyword queries
- [ ] Run OpenAlex `works?search=` (+ year/OA filters when needed); use `mailto=`
- [ ] Add English terms for international packs
- [ ] Score & keep Top 50
- [ ] Prefer Pack A/H/A-CN primary (OpenAlex DOI/OA · arXiv/PDF) over media blogs

FilesExpand file tree

sources-reference.md

Latest commit

History