• Pass just single request with all the context, including bar plots and tables.
• Make it more stable. Why temperature=0.0 seems not to work? Use structured outputs.
• Support other plot types.
• Two-step approach: summary for each plot, then overall summary based on it.
• Separately generate recommendations that users can expand on click.
• Try with o1-preview.
• Try with not passing full column descriptions, but instead use short ones, and a separate table expanding the detail.
• Support Anthropic
• Support Seqera Chat
• Add "Continue in Seqera Chat button"
• Generate markdown, then convert to HTML
• Use a backend endpoint to create user session with the question/response
• Parse sample names and add links to highlight those on the report
• Add "Show More" button
• Streaming
• Calling completions from FE
• Configuration modal in FE to select provider, API key, model name
• Short and full summary, and "generate more" in FE
• Use the same data in FE and BE generations, which will save on passing the base64 data from BE
• Match python generated report data with JS
  • JS report content: support multiple datasets
  • PY report content: all plot types
  • PY report content: nice naming of plot types - same as in JS
  • PY report content: suffixes in line plots and other plots
  • PY report content: separate violin and table representations? Or make them the same in JS?
• Add general stats dynamic summary
• Handle button name when table switch to violin and back
• Handle when prompt is too long (PY)
• Handle when prompt is too long (JS)
• Add command line options
• Seqera AI support
  • Support new endpoint https://github.com/seqeralabs/web/pull/204
  • Auth: read API token from Platform environment?
  • Support sign in into Platform?
• Handle when report was run without --ai-summary flag:
  • show all buttons next to plots from the toolbox
  • toolbox button to run AI summary on the entire report
  • add "Copy for AI prompt" button on the top level summary
• Handle very large reports:
  • For large reports, use only general stats
  • For large tables, use the reduced violin data dicts
  • If table is large, do not use hidden columns
  • Support in PY
  • Support in JS
• JS: exclude hidden samples from prompt
• Toolbox:
  • Dynamic info based on provider
  • Links to tokens for Seqera AI
  • Links to available models
  • Models autosuggest
• Documentation for AI features
• Track basic usage
• Large reports: allow to hide some metrics and retry (fix modal for violin plots)
• Be helpful when user hidden all columns and/or samples
• Draw a new "Copy for AI" icon - a clipboard icon combined with a sparkle?

Ideas to handle large reports

1. Define token limits for each provider based on their documented context windows

2. Prioritize content:

   • Always include the tools summary for context
   • Always include general statistics if available
   • Add detailed sections until approaching the limit

3. Use a conservative token estimation (4 chars per token) and leave a 10% buffer

4. Log a warning when content needs to be truncated

Additional recommendations that could be implemented:

1. Add configuration options to control content prioritization:

config.ai_summary_prioritize_sections = ["fastqc", "alignment_metrics"]  # Priority sections
config.ai_summary_exclude_sections = ["software_versions"]  # Sections to skip

2. Implement smarter content selection:

• Skip redundant information
• Summarize large tables/plots
• Focus on sections with warning/error flags

3. Consider splitting analysis into multiple API calls for very large reports:

• First call for high-level overview
• Additional calls for detailed section analysis
• Combine results intelligently

4. Add provider-specific optimizations:

• Use different prompts/strategies based on model capabilities
• Leverage features like function calling for structured output