More context feels safer. Include the entire conversation. Retrieve ten documents instead of three. Add everything the user has ever done.

Then the answers get slower, more expensive, and sometimes worse.

In this session, we will run a controlled experiment against a support assistant. We will ask the same questions while changing only the context strategy: full conversation history, fixed-size retrieval, reranked retrieval, summaries, and explicit application state.

For each version, we will measure answer quality, groundedness, latency, and token usage. Then we will deliberately introduce irrelevant documents, stale memory, and contradictory instructions to see where the system fails.

The demo will show how to:

- Build a small evaluation dataset for context decisions
- Separate conversation history, retrieved knowledge, and application state
- Detect when additional context reduces answer quality
- Compare fixed retrieval with filtering and reranking
- Summarize older history without losing critical facts
- Exclude sensitive or irrelevant information by design
- Create a context budget based on evidence instead of intuition

This is not a session about finding the perfect context-window size. It is about treating context as an engineered input that can be tested, measured, and constrained.

Attendees will leave with a repeatable experiment for deciding what their AI system needs to know—and what it should never receive.