Goal

Use this when the user gives a concrete topic, method, author, lab, or venue.

python3 scripts/semantic_scholar_daily.py search \
  --query "{user query}" \
  --days 180 \
  --limit 10 \
  --output-json semantic_scholar_search.json

Optional filters:

• --author
• --venue
• --config with --domain

After running, read the JSON and present a concise ranked list with title, date, venue, citations, and one-line relevance.

Mode 2: recommendation ranking

Use this when the user asks for latest papers in their field, a daily digest, or paper recommendations.

python3 scripts/semantic_scholar_daily.py recommend \
  --config "$OBSIDIAN_VAULT_PATH/99_System/Config/research_interests.yaml" \
  --vault "$OBSIDIAN_VAULT_PATH" \
  --days 45 \
  --hot-days 365 \
  --recent-cutoff-days 30 \
  --per-domain 20 \
  --top-n 10 \
  --max-per-domain 3 \
  --output-json semantic_scholar_daily.json

What this does:

• builds a recent candidate pool from the last --days
• builds a hot backfill pool from the older --hot-days range, excluding the newest --recent-cutoff-days
• merges and reranks both pools
• applies a per-domain cap so one topic does not dominate the final top 10
• reuses existing paper notes when possible

If the user gives a target date, add --target-date YYYY-MM-DD.

Mode 3: write the Obsidian note

Use this after recommend whenever the user asks to save or refresh the daily paper note.

python3 scripts/semantic_scholar_daily.py write-note \
  --input-json semantic_scholar_daily.json \
  --output-md "$OBSIDIAN_VAULT_PATH/10_Daily/YYYY-MM-DD论文推荐.md" \
  --vault "$OBSIDIAN_VAULT_PATH" \
  --link-keywords

What this does:

• writes the final markdown note in the required Obsidian layout
• adds keywords and tags frontmatter
• links existing paper notes in the body
• optionally auto-links note keywords from 20_Research/Papers/ into the generated markdown

Workflow

1. Read the research-interest config.
2. Run search for direct topic lookup, or recommend for a daily digest.
3. For daily recommendations, keep exactly 10 papers unless the API returns fewer valid candidates.
4. Prefer the script's reranked output; it already mixes recent and hot candidate pools.
5. Check existing notes before expanding any paper entry.
6. If the user wants a saved note, run write-note.
7. For the top 3 papers, do follow-up manually in the agent when the user clearly wants deeper work:
   • search for an existing note by paper id, exact title, and distinctive title keywords
   • if a note exists, reuse that wikilink
   • if no note exists, use extract-paper-images and paper-analyze
8. In chat, summarize the trends and point to the saved note instead of dumping the markdown.

Ranking expectations

The recommendation script combines:

• relevance_score: match to research interests
• recency_score: stronger boost for recent papers
• popularity_score: citation and early-attention signals
• quality_score: abstract-based technical contribution estimate
• diversity control: --max-per-domain limits single-domain domination

Interpret the final list as a curated reading queue, not a raw search dump.

Note structure

The markdown note should contain:

---