Context Quality Engineering
Context quality engineering is the production discipline of selecting, compressing, deduplicating, and structuring only the decision-relevant information an LLM or agent needs.
Key points
- Datadog reports that model context windows have expanded from roughly 128,000 tokens to as high as two million tokens in some pricing tiers [src-037].
- Larger windows have shifted the constraint: the problem is often no longer raw capacity, but whether the inserted context actually helps the model decide [src-037].
- Datadog observed that average request tokens more than doubled for median customers and quadrupled for 90th-percentile power users year over year [src-037].
- As teams stuff conversation history, retrieved documents, tool outputs, state, and guardrails into prompts, noise and redundancy can bury critical details [src-037].
- The report states the new limiting factor clearly: context quality, not context volume, determines reliable production agents [src-037].
- Effective systems maintain retrieval quality, summarization, deduplication, compression, and information hierarchy so long-context models receive high-signal inputs [src-037].
- Google Cloud adds a governance-related quality constraint: policy text can be necessary, but too much policy in-context can distract the model from the task; hard guardrails can keep policy outside the reasoning window [src-043].
- [[Context-sharding]] keeps each agent focused on a smaller work slice with only the instructions and skills needed for that slice [src-043].
- Knowledge Catalog and Workspace Intelligence represent a productized version of context quality: organize data semantics and workflow context so agents receive trusted, relevant business context instead of unstructured dumps [src-044].
Related entities
Related concepts
- Context Engineering
- Context Rot
- Retrieval-Augmented Generation (RAG)
- LLM Knowledge Bases (Karpathy pattern)
- Prompt Caching for Agents
- Claude Code Token Economics
- Context Sharding
- Containment Over Constraint
- Enterprise Knowledge Graph
- Workspace Intelligence