Add one-shot prompt generator for external LLM figure analysis

[amahpour]

Summary

This PR introduces an automated prompt generation system that creates self-contained markdown files for external LLM analysis of low-confidence figures. The system collects figures flagged by the local processor and organizes them into batches with explicit instructions for parallel processing and status reporting.

Key Changes

• New module src/prompt_generator.py: Generates one-shot markdown prompts containing:

  • All figures below the confidence threshold, grouped into batches of 20
  • Absolute image paths for direct use with vision-capable LLMs (Claude, Codex, etc.)
  • Instructions for parallel processing and batch-level status updates
  • Optional integration with analysis schema templates
  • Output paths for saving results back to the analysis directory

• Refactored confidence scoring in src/local_processor.py:

  • Replaced binary SIMPLE_TYPES/COMPLEX_TYPES classification logic with a continuous _compute_confidence() function
  • Introduced CONFIDENCE_THRESHOLD (0.7) to determine which figures need external analysis
  • Confidence scoring now considers description length and classification specificity
  • Updated docstrings to clarify the distinction between local resolution and external queuing

• Integrated prompt generation into pipeline (src/pipeline.py):

  • Added call to write_prompt() after local figure processing completes
  • Generates external_analysis_prompt.md in the PDF output directory
  • Logs the prompt path for user reference

• Updated documentation (prompts/figure_analysis.md):

  • Simplified batch usage instructions to reference auto-generated prompts
  • Removed manual workflow details in favor of pipeline-generated prompts

• Comprehensive test coverage (tests/test_prompt_generator.py):

  • Tests for empty figure sets, batching logic, filtering, and file I/O
  • Validates inclusion of absolute paths, subagent instructions, and template integration

Implementation Details

• Figures are loaded from processing/fig_*.json status files and filtered by status == "needs_external"
• Batches are created with explicit size limits to enable parallel subagent processing
• The prompt includes Moondream's local classification and confidence scores for context
• Output paths are pre-computed so external LLMs can save results directly to the analysis directory structure
• The system gracefully handles cases where no figures need external analysis (returns None)

https://claude.ai/code/session_0139BXTEb42aXeukRzsaxyEQ