Hi everyone. I’ve been testing DeepSeek Chat for some technical design sessions around migrating an old on-premise monolith to Azure. For the first 15-20 exchanges, it’s incredibly helpful—clear, step-by-step breakdowns, good Azure service recommendations. But then, almost like hitting a wall, the output quality seems to fall off a cliff.
Around that 20-message mark in a single conversation, the responses become noticeably less detailed. It starts repeating points it already made earlier, or gives very generic advice like “consider refactoring the database” without the earlier nuance. It feels like it’s losing the thread of my specific architecture, even though I’m referencing our previous discussion.
Has anyone else run into this? I’m nervous because I was hoping to use longer sessions to work through entire migration phases, like nailing down the database migration strategy from SQL Server to Azure SQL. But if the context window gets fuzzy or the model’s focus drifts that quickly, it’s hard to rely on it for a full planning workflow.
Is this a known behavior? Are there any tricks to keep the output consistent in longer chats, besides just starting a fresh thread? Really hoping there’s a way to manage this, as the initial advice is so good.
One step at a time
Yeah, I see this too. It's not just you. The context window is sliding, not static, so earlier details get dropped even if you're referencing them.
I've worked around it by breaking the session into phases. Start a new thread for each major phase, like "Phase 1: Database Migration Strategy". Paste the key architectural decisions from the previous chat as your new first message to re-establish context. It's annoying but keeps the quality up.
For your SQL Server to Azure SQL migration, you could have one thread for assessment, another for the actual cutover steps, and a third for post-migration tuning.