Skip to content

feat: 完成 M05 论文检索、相关性筛选与 OA 全文本地化 - #7

Open
wdwd200 wants to merge 2 commits into
redmaplewww:mainfrom
wdwd200:feat/m05-relevance-selection
Open

wdwd200 wants to merge 2 commits into
redmaplewww:mainfrom
wdwd200:feat/m05-relevance-selection

Conversation

@wdwd200

@wdwd200 wdwd200 commented Aug 24, 2026

Copy link
Copy Markdown

Summary

完成 M05 论文检索第一版的端到端实现,并整合此前的 OA full-text discovery 工作。

主要包括:

  • 基于 LLM 将原始研究问题改写为 3 条互补的学术检索 query
  • 使用 OpenAlex、Crossref、Semantic Scholar 进行真实多源论文检索
  • DOI / 标题去重
  • 对全部去重候选进行批量 LLM 语义相关性筛选
  • 使用原始 research idea 作为相关性判断目标
  • Reader 当前工作集最多 20 篇
  • 显式 user seed papers 优先
  • LLM 不可用时使用 lexical relevance 排序作为 fail-open fallback
  • relevant overflow papers 作为 deferred candidates 保留,而非判为不相关
  • Unpaywall + OpenAlex OA discovery
  • 合法 OA PDF 下载并写入 PaperRecord.full_text_path
  • Reader 可直接消费本地 PDF
  • 增加 opt-in live M05 smoke test 工具
  • 不绕过付费墙,不抓取 OA landing page HTML

Live verification

已进行了真实环境 smoke test。

OA-only

使用真实 DOI:

10.3389/frai.2024.1293084

结果:

  • Unpaywall lookup succeeded
  • OA status: open_access
  • provider: unpaywall
  • license: cc-by
  • real PDF downloaded locally
  • PDF size: 733733 bytes
  • PaperReaderService successfully consumed the PDF
  • Reader locator count: 10

Full M05 E2E

Research idea:

How can retrieval augmented generation reduce hallucinations in large language models?

结果:

  • real OpenAI-compatible LLM query rewrite succeeded
  • exactly 3 scholarly retrieval queries generated
  • real multi-source scholarly search executed
  • 17 observed/deduplicated candidates
  • 10 papers entered the Reader workset
  • semantic LLM batch failures: 0
  • 8 papers resolved as open access
  • 2 had no open full text found
  • 4 real PDFs were materialized locally
  • 4 local PDFs were successfully consumed by Reader
  • Reader failures: 0
  • full search-to-Reader chain completed successfully

No credentials or email values are stored in the repository.

Known limitation

OPEN_ACCESS does not always imply that the provider returns a directly downloadable PDF.

Some legitimate OA records resolve to HTML landing pages. The first version intentionally rejects HTML instead of scraping publisher pages or bypassing access controls.

Verification

  • Focused M05 tests: 35 passed
  • Full pytest: 53 passed
  • Strict pytest: 53 passed
  • Ruff: passed
  • autoresearch doctor: healthy, exactly five agents
  • .ai-team/check.mjs --base upstream/main: valid, 8/8
  • git diff --check: passed

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant