08 / Exploration
Agentic demos for Algolia
Three article companions for Algolia: an adaptive-ranking comparison, a lexical-versus-semantic search benchmark, and a measurement surface for the cost and value of an AI agent.
Context
An article can explain the idea. A demo has to let someone poke it and see where the idea holds—or stops holding. These companions keep the inputs visible and the claims small.
The problem
Agentic search is easy to describe vaguely. The demos needed to make ranking behavior, retrieval tradeoffs, agent value, cost, latency, uncertainty, and failure states inspectable without pretending fixture data is production evidence.
Work involved
I designed and built these companion demos in Next.js and TypeScript, including the fixture models, evaluation logic, interactive surfaces, and the explanatory boundaries around each result.
Approach
- Built an Adaptive Intent comparison showing standard ranking, behavioral ranking, evidence, and fallback when the fixture does not qualify.
- Built a Vector Search Revisited benchmark comparing lexical, semantic, and hybrid retrieval with judged queries, hard filters, and an exposed blend weight.
- Built an Agent Value Measurement surface comparing control and agent exposure across cost, latency, uncertainty, and segment behavior.
- Kept the examples deterministic and labeled the data as synthetic or fixture-level, because a tidy demo is not production evidence.
What this work shows
The demos do not ask the reader to trust a tagline. They expose the inputs, ranking or measurement logic, controls, caveats, and fallback states so the argument can be checked.
Technical or communication skills involved
Outcome
Three article companions that make agentic search and AI-agent tradeoffs easier to inspect: one ranking comparison, one retrieval benchmark, and one value-measurement surface.