Summary

• Implemented a custom search with Exa
• Together with Gemini, this allows for good search grounding across our categories
• Costs Estimation: ~160$ for 1,000 entries
  • The costs for the research are currently ~8ct per entry
  • The total costs for an entry are therefore ~16ct
  • For 1000 pages in our glossary, this means $160 spend

Insights

Demo

For the entry single-sign-on, we now get a list of 26 sources, which looked all like pretty good resources to me:

Before we only had the Google results (and only the top 3), so I do think this is an improvement in quality.

The Implementation

Why Exa (vs Deep Research)?

Next Steps

• Finish the research branch (Monday)
  • Clean up the feature branch (make it ready to merge & deploy on Monday)
  • Add Search Caching in DB to avoid duplicate exa requests
• Connect our 4 generate workflows (takeaways, content, FAQs & metadata) to the research task (Friday)
• Update create_pr workflow to allows for field-level updates (next week)

Open Questions

• LLM Observability: Any tools in use/experienced to share?