Your assistant answers the first turn well and the third turn badly. Users say it forgets. What do you check first?

A strong diagnostic approach involves inspecting the input at the retrieval stage to see if the query has been corrupted or diluted. Often, the issue is that the current question is lost within a massive concatenation of previous turns, or that the query lacks necessary context (like an antecedent for a pronoun) because of how the conversation history was pooled. Focusing on the retriever input allows you to identify if the fault lies in context management or query construction, rather than assuming the model’s generation is the primary failure point.