Developers Urged to Give AI Agents 3rd-Stage Search Tools Beyond 2024 Vector Retrieval
Updated
Updated · The New Stack · Jun 21
Developers Urged to Give AI Agents 3rd-Stage Search Tools Beyond 2024 Vector Retrieval
3 articles · Updated · The New Stack · Jun 21
Summary
Perplexity’s “search as code” push is framed as a third stage in AI retrieval, with developers urged to let agents use professional-style search controls rather than consumer search defaults.
2024’s vector-database approach improved little because isolated text chunks and similarity scoring often missed context, while hybrid methods such as BM25 and learned ranking moved more use cases into production.
AI agents can exploit richer tools that humans rarely use—proximity search, semantic search, date filtering, grouping and aggregation—then chain those queries to build overviews, test hypotheses and verify details.
The article argues implementation is already straightforward: models can write complex queries if developers describe available fields, filters and ranking options in plain text.
The broader shift is from optimizing search for casual users to exposing a wider retrieval toolbox for agents, more like systems built for professional researchers or quants.