Search Knowledge

• Keywords: Technical terms, method names, model names (e.g., "MCP", "RLHF", "attention")
• Authors: Person names (e.g., "Danqi Chen", "Ashish Vaswani")
• Domains: Research area names (e.g., "agent", "safety", "reasoning")
• Concepts: Broader ideas (e.g., "test-time compute", "tool use", "injection")
• Modifiers (optional):
  • - tag: TAG — only search notes with this tag in the filename
  • - type: notes|syntheses|daily|all — which knowledge sources to search (default: all)
  • - verbose — show full matching context instead of summaries

Step 2: Collect Knowledge Sources

Determine which sources to search based on - type modifier:

# Notes (default: always)
find knowledge/ -name "summary_*.md" -not -path "knowledge/syntheses/*" | sort

# Syntheses
find knowledge/syntheses/ -name "*.md" | sort

# Daily lists
find knowledge/daily/ -name "*.md" | sort

If - tag: TAG is specified, only include notes matching summary_{TAG}.md.

If no sources exist, report that the knowledge base is empty and suggest running /daily-papers and /read-paper first.

Step 3: Search Across All Sources

Use Grep to search for the query terms across all collected sources. Search with multiple strategies:

1. Exact keyword search: Grep for the exact query terms
2. Related term search: Grep for semantically related terms (e.g., "injection" → also search "adversarial", "attack", "security")
3. Author search: If the query contains a name, search for it in author fields

For each match, capture:

• Source file: Which note/synthesis/daily list
• Paper title: From the note header
• Match context: The surrounding 2-3 lines showing the match
• Section: Which section of the note the match appears in (Problem, Method, Key Insights, Connections, etc.)

Step 4: Organize and Rank Results

Group results by relevance:

1. Direct matches — query terms appear in title or core method description
2. Method matches — query terms appear in Method or Key Equations sections
3. Insight matches — query terms appear in Key Insights or Connections sections
4. Mention matches — query terms appear in passing (Questions, Connections)

Step 5: Present Results

Present results in a structured format:

Found {N} matches for "{query}" across {M} papers

## Papers where "{query}" is central

### {Paper Title} ({arXiv ID})
- **Tag**: {tag}
- **Why relevant**: {one-line explanation of the match}
- **Key finding**: {the most relevant insight or method detail}
- **Note**: knowledge/summary_{tag}.md

## Papers mentioning "{query}"

### {Paper Title} ({arXiv ID})
- **Context**: {where and how it mentions the query}
- **Note**: knowledge/summary_{tag}.md

## Synthesis insights

From synthesis_{date}.md:
- {relevant cross-paper finding or trend}

## Daily paper matches

From {date}.md:
- {paper title} — {tier} — {why relevant}

If - verbose is specified, include the full matching context for each result.

If no matches found, suggest:

• Alternative search terms
• Running /daily-papers to discover new papers on the topic
• Running /read-paper on specific papers to add them to the knowledge base

Step 6: Suggest Follow-ups

Based on search results, suggest:

• Papers that could be read next (/read-paper ID)
• Running /synthesize - full if the knowledge base has grown significantly
• Running /daily-papers to find more papers on underrepresented topics

Key Rules

• PDF IS BANNED. Never download or link to PDF files.
• Search is local only. Only search within knowledge/ directory contents.
• Be thorough. Use multiple search strategies (exact, related, author) to avoid missing relevant papers.
• Rank honestly. Don't inflate relevance — distinguish between "central to the paper" and "mentioned in passing."
• Preserve context. Always show enough surrounding text for the user to understand why a match is relevant.
• Report empty results clearly. If nothing matches, say so and suggest alternatives rather than returning weak matches.