5.4R · Relationship●●○○○
Context rot: when more context hurts
Past a point, extra context backfires: the key fact gets buried, attention spreads thin, latency and cost rise, and answers get worse. Curate, compress, retrieve the few right chunks, cite, and verify — don't dump.
You'll get more from this if you've seen4.2Context window & KV cache5.2RAG: retrieval as a callback to similarity
1Predict first: does dumping in more text help it find the answer?your turn
The wall
The model reads everything you give it. So the safe move is to paste in all your docs and let it sort them out — right?
You have one short fact that answers the question. Does pasting 10,000 tokens of unrelated text around it help the model find that answer?
space play/pause←→ stepR replay
Common questions
What is "Context rot: when more context hurts" about?
Past a point, extra context backfires: the key fact gets buried, attention spreads thin, latency and cost rise, and answers get worse. Curate, compress, retrieve the few right chunks, cite, and verify — don't dump.
What problem does it solve?
The model reads everything you give it — so just paste in all your docs and let it sort them out, right?
What will I be able to do after this lesson?
You can explain context rot: why more context can lower answer quality, and the habits (curate, compress, retrieve, cite, verify) that beat dumping everything in.
What comes next?
Deciding what the model should see is real work — next, the tools and loops that manage it for you.
Deciding what the model should see is real work — next, the tools and loops that manage it for you.
6.1 The harness: the loop, made real