Beyond embeddings: grep agents and whole-corpus long-context prompts as underrated retrieval patterns

jakedahn · x · 2026-09-25

The author lays out three patterns for agents to find context:

Embedding clusters: using the centroid of a cluster of interesting things as a coarse filter, though it can be noisy and misleading.
Agent-driven grep: letting an agent literally grep through files until satisfied, then feeding results as next-turn context — works well in practice and mirrors how programmers work without agents.
Whole-corpus in a giant context window: cat everything into a wall of text and prompt for "helpful context for fixing this bug," "find the interestingness," or "find the incidental vs essential complexity" — the last being a promising proxy for extracting rigid APIs from slop code.

Doing this purely with LLMs is expensive; models like jev make surfing at this layer fast and cheap.

More from coding & agent