Louis-François Bouchard, Omar Solano and Samridhi Vaid examine context engineeringContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. through a course-grounded AI tutor. The workshop distinguishes managing one conversation from remembering information across sessions, then tests whether compression, retrieval and caching actually improve the tutor's answers, cost and latency. Its findings concern the team's workloads and evaluation design rather than a universal rule for every agent.
Louis-François Bouchard explains that instructions, conversation history and retrieved tool outputs compete for context space. He reviews trimming, selective retention, summarization and progressive disclosure, but emphasizes that rewriting an otherwise reusable prompt can invalidate cached inputPrompt caching reuses compatible computation from previously processed prompt content to reduce repeated model-input work.. The relevant optimization is therefore the useful information available per request and its actual cost, rather than simply sending fewer tokens.
Omar Solano describes a single-agent tutor with hybrid semantic and keyword retrievalRetrieval is the process of selecting relevant stored information and returning it to an AI system for the current task., reranking and course-source filtering. A separate read-only knowledge-base browsing tool adds filesystem exploration under command, output and sandbox limits. On the tested student questions, that extra browsing did not improve retrieval recall and added latency, illustrating why a more elaborate retrieval design needs workload-specific evidence.
Omar Solano evaluates real student question-and-answer pairsEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. alongside synthetic multi-turn fact-recall sessions, logging token usage, cache hits, tool calls, cost and time to first token. Keeping history intact outperformed the initial clearing and summarization defaults in those tests, partly because clearing tool outputs forced repeated retrieval. The initial sessions were short and used few trials, so the reported ordering should not be treated as a broadly established benchmark.
Samridhi Vaid extends the experiments to a cheaper hosted model, longer fact-recall probes, document retrieval and a hardware-constrained local setup. Cached full history preserved details better in the tested hosted chats, while retrieval became necessary when documents exceeded the local context windowA context window is the maximum amount of tokenized information an AI model can consider during one processing session.. The final tutor choice combines DeepSeek V4 Flash, hybrid retrieval and preserved history up to a 30,000-token compaction threshold. The practical conclusion is to identify the binding constraint and measure the tradeoff, rather than compact every conversation by default.
Watch on YouTube
