pdf-processing

PDF Processing Skill

When This Skill Activates

• When PDFs need to be indexed for semantic search
• When importing technical documentation or research papers
• When adding documents to the knowledge base
• When user wants to query PDF content semantically

Quick Path (Single PDF — No Agent Needed)

For a single PDF, run directly without spawning an agent:

# Index a PDF into T3 (docs__ collection)
nx index pdf /path/to/file.pdf --corpus {corpus-name} --monitor

# Index into knowledge__ collection (for reference material)
nx index pdf /path/to/file.pdf --collection knowledge__{name} --monitor

# Dry-run preview (extract and embed locally without storing)
nx index pdf /path/to/file.pdf --corpus {corpus-name} --dry-run

# Force re-index (bypass staleness check)
nx index pdf /path/to/file.pdf --corpus {corpus-name} --force --monitor

# Verify indexing — use search tool: query="representative query", corpus="docs__{corpus-name}", limit=3

Corpus naming: Use author-year-short-title pattern. The --corpus flag auto-prepends docs__ — do NOT include the prefix.

Agent Invocation (Batch/Complex Scenarios)

Delegates to the pdf-chromadb-processor agent (haiku) for:

• Multiple PDFs needing batch processing
• PDFs from URLs (download + index)
• Complex corpus organization decisions

## Relay: pdf-chromadb-processor

**Task**: [what needs to be done]
**Bead**: [ID] or 'none'

### Input Artifacts
- Files: [PDF paths or URLs]

### Deliverable
Indexed PDF content in T3 via `nx index pdf`, with chunk counts and searchability verified

### Quality Criteria
- [ ] All PDFs indexed via `nx index pdf`
- [ ] Content searchable via search tool
- [ ] Processing report with chunk counts

For full relay structure and optional fields, see RELAY_TEMPLATE.md.

Success Criteria

• All PDFs indexed via nx index pdf (not manual store_put tool)
• Corpus names follow author-year-short-title convention
• Sample queries return relevant results via search tool
• No sandbox permission issues (all extraction runs inside nx index pdf native pipeline)

Agent-Specific PRODUCE

Outputs generated by the pdf-chromadb-processor agent:

• T3 knowledge: Indexed PDF content via nx index pdf pipeline (Docling extraction, context-safe chunking, Voyage embeddings, atomic storage)
• T2 memory: Indexing status log via memory_put tool: content="content", project="{project}", title="pdf-index-log.md", ttl="30d"
• T1 scratch: Working notes during processing via scratch tool: action="put", content="Processing: {filename} - {N} chunks indexed", tags="pdf-processing"

Session Scratch (T1): Agent uses scratch tool for ephemeral working notes during the session. Flagged items auto-promote to T2 at session end.