Gabriele Farei@jayfarei
Last Saturday I joined the @cerebral_valley x @MongoDB AI Hackathon in London 🙏
I explored adaptive retrieval: how can a data interface improve every time an agent uses it?
The prototype I built was:
1/ Mount a dataset interface as a virtual filesystem, so usage memory lives next to the data surface.
2/ Let agents use bash to search, sample, inspect schemas, and plan trajectories.
3/ Commit the final trajectory as an executable TypeScript file.
4/ Let that file compose deterministic primitives with skill-driven agent steps via @FredKSchott's Flue ❤️
The challenge was to force claude/codex to "exernalise" the trajectory, I did that by offering only sample data to derive functions and skills, but forcing the execute() step to be a single typescript file.
The result is that over time, repeated queries crystallise into intent-shaped interfaces. More and more follow-on queries were simply repeats of the previous ones with different arguments, leading to fewer exploratory steps, more direct calls, faster answers.
The useful win win of this approach:
1/ data providers get a concrete signal for how agents actually want to use their data,
2/ they can expose cheaper interfaces based on derivative usage.
3/ they can implement surgical optimisations to existing user trajectories (replacing expensive searches, LLMs steps that could be deterministic, or cheaper models, ...)
Cleaning up the project to release it soon 👉 datafetch.ai
