A strong context sample is a curated, compact set of information (text, memory, retrieved docs) that gives the model what it needs to reason, without overwhelming it. First, you pick the key elements: the user query, relevant retrieved documents or knowledge bits, memory or history summaries, and system instructions. Choose only the most semantically relevant pieces. Use embedding similarity, metadata filters (e.g. recency, importance), or heuristics to select top-k candidates.
Next, you format them clearly and consistently. Segment sections (for example: “History: …”, “Relevant Docs: …”, “Current Query:” ), so the model can distinguish sources. You might also include separators or labels. If content is long, you might truncate or summarize documents. In long conversations, compress past turns into short summaries to retain coherence while conserving token space.
Finally, you test and iterate. Run the model with that sample context, see errors or hallucination, remove or adjust context pieces, and see if performance improves. You might compare variants (different context subsets) to evaluate which helps most. Over time, the “good context sample” becomes a template or pattern your system reuses with different inputs.