Context Filter
Filters a big tool output before an agent's LLM reads it (context pollution). A GitHub issues list (30 issues) or a CI log (26 lines) is split into items in code, and Jev judges each item on its own, all in parallel: is this item needed to answer the question? Only the kept items reach the answer. Against it: sending the whole output, an LLM that summarizes first, and a keyword filter (grep, $0). The same model answers in every variant, and Jev's grader scores every answer. Traced. Measured: pass rate, cost per passing answer, tokens given to the answer, and item precision and recall.
With Jev
◐ ▮▮▮ ▦Without Jev
LLM · CODECalls Jev and every baseline for real, traced in Langfuse. Nothing is saved.
Runs every labeled row, one at a time: 14 rows × every variant, all real, paid calls. Rows run sequentially so latency measures the model, not a traffic jam. The run is saved as it goes and lands in History.