What I learned making an evidence-first local search layer for coding agents
retained text
I kept seeing the same failure mode: the agent found a plausible file, answered confidently, and never proved that the excerpt still matched the file on disk. I built sift-light around the opposite workflow: exact search first, explicit semantic candidates when opted in, and a follow-up that can inspect and validate the retained source. The recent release also adds visual retrieval for images and scanned PDF pages on Apple Silicon, while keeping the original page number and source hash around. The hard parts were less glamorous than the model: - keeping literal evidence from being displaced by semantic candidates - making pagination and inspection reuse one snapshot instead of rerunning a query - rejecting a remote media response that names a file the local admission layer never sent - bounding Base64 request size before reading large files - making stale source evidence visible after edits I’m curious how other agent builders handle this. Do you keep a strict distinction between “relevant” and “verified,” or do you let the model decide when a candidate is good enough?
document fetched 2026-10-13 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1x3z7ki/what_i_learned_making_an_evidencefirst_local