01 / FRESH CONTEXTRetrieve for the question, not just recency
Begin with a question class rather than a growing pile of recent items. A version-comparison question needs different evidence from a current-headlines question. Use source, version, language, and time constraints where they matter. Inspect the retrieved passages before judging the final answer; fluent wording can hide an irrelevant candidate set.
02 / FRESH CONTEXTCarry source identifiers into every chunk
Every passage should remain connected to its source item and revision after chunking. Keep publication and observation times distinct. When the source changes, those relationships let you find affected chunks and review answers built from earlier evidence. A document fragment without a source trail is difficult to cite or invalidate responsibly.
03 / FRESH CONTEXTTest citations and abstention as well as fluency
Test questions that have an answer, conflicting evidence, and no answer in the collection. Check that citations resolve to supplied sources and support the nearby claims. Vary distractors and context placement under your own conditions. A declared input capacity is not a substitute for observing how the complete pipeline performs on your task.
Answer a small set of questions well
Select a modest source collection and write questions that require different passages within it. Add a question that the collection cannot answer. Record which evidence should support each response, then inspect retrieval and generation separately. This gives you a practical baseline for chunking, freshness, and citation checks before a larger corpus makes the source of an error difficult to identify.