CORE GUIDE
Relevance over raw volume
More input can reduce decision quality when low-value material competes with the evidence that actually matters.
Mental model
Context quality is a selection problem: include the smallest set of information that preserves the evidence and constraints required for the next decision.
Why it matters
Teams often react to model mistakes by adding more documents, more examples and more history. That can increase noise, contradiction and cost. Strong systems rank information by task relevance, authority and recency instead of treating context as a dump.
01
Selection before generation
The application assembles context from multiple sources. Each candidate item has a value for the current decision and a cost in tokens and attention. Retrieval, filtering and summarization are therefore part of reasoning quality, not merely infrastructure optimizations.
02
Example: policy answer with fifty documents
A support assistant receives forty old FAQ pages and one current policy. The current policy contains the answer, but its signal is diluted by contradictory legacy text. Filtering by authority and effective date produces a smaller context and a more reliable answer.
Common failure modes
- Appending every retrieved result to the prompt.
- Optimizing only semantic similarity while ignoring authority.
- Keeping background information that is unrelated to the current decision.
Engineering heuristics
- Rank context by relevance, authority and freshness together.
- Set explicit caps per evidence class instead of one global token limit.
- Log what was excluded so retrieval mistakes remain diagnosable.
Takeaways
- 01Useful context is curated, not accumulated.
- 02Smaller can be better when it preserves the right evidence.
- 03Context selection is part of application correctness.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.