Helping Claude Code to find undocumented DuckDB APIs from the code (flaky.build)

A developer has successfully leveraged Claude Code, an AI coding assistant, in conjunction with GitHits, a code search tool, to discover undocumented DuckDB C++ APIs critical for enhancing a web archive extension, duckdb-web-archive. While Claude Code initially struggled, confidently asserting that access to certain capabilities was impossible, GitHits allowed for real-time searches across actual code repositories. This enabled the developer to implement predicate pushdowns, which optimize data fetching efficiency in SQL queries by filtering and limiting data at the API level rather than locally. The integration of GitHits proved transformative, as it revealed concrete examples of function signatures and implementation patterns from existing DuckDB extensions like httpfs and postgres_scanner. Consequently, the developer completed the extension with full predicate pushdown support, drastically reducing the number of HTTP requests required for queries. This case highlights the significance of combining AI tools with comprehensive code search capabilities, addressing the challenge of undocumented APIs in complex programming environments, and showcasing how such integrations can enhance developer productivity in the AI/ML community.