--- title: Context Engineering in 2026 — Louis-François Bouchard,... description: A workshop from Towards AI walks through real production experiments on managing LLM context windows for an AI tutor product used by their students. The team... canonical: https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Context Engineering in 2026 — Louis-François Bouchard, Omar Solano & Samridhi Vaid, Towards AI | daily.dev og:description: A workshop from Towards AI walks through real production experiments on managing LLM context windows for an AI tutor product used by their students. The team... og:url: https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog og:image: https://api.daily.dev/og/posts/pSjpEiNoG.png og:image:alt: Context Engineering in 2026 — Louis-François Bouchard, Omar Solano & Samridhi Vaid, Towards AI og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Context Engineering in 2026 — Louis-François Bouchard, Omar Solano & Samridhi Vaid, Towards AI **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 63 min read · 0 upvotes · 0 comments ## Summary A workshop from Towards AI walks through real production experiments on managing LLM context windows for an AI tutor product used by their students. The team tested compaction strategies (sliding window, summarization, selective retention, prompt compression) against simply keeping full chat history, across Gemini 3.5 Flash and DeepSeek v4 Flash, plus a local MacBook setup. The consistent, counterintuitive finding: keeping full history outperformed compaction on cost, latency, and memory recall, because prompt caching makes resending the same tokens far cheaper (up to 50x with DeepSeek) while compaction breaks the cache and forces full-price recomputation. They also compared RAG vs GraphRAG (RAG won on cost/results), tested letting the agent browse a wiki-style knowledge base via bash commands (no recall improvement, 50% slower), and found hybrid search (dense + BM25) necessary since pure semantic search collapsed to 0% recall at 400k tokens. They ultimately shipped DeepSeek v4 Flash with hybrid retrieval and a 30k-token compaction threshold, only compacting when truly forced by context window limits (e.g., local models capped at 32k). ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does compacting or summarizing chat history actually save money when using LLM prompt caching? No, in most cases keeping the full chat history is cheaper than compacting. Summarizing or truncating context breaks the provider's prompt cache because it creates a new context, forcing the model to pay full price for those tokens again. Since cached tokens can cost up to 50 times less on some APIs like DeepSeek, compaction only pays off if it compresses the context by more than 50x without losing quality. _Anyone tuning agent costs around prompt caching can find hands-on breakdowns like this through daily.dev._ ### Does using RAG or full chat history give better recall than summarization in a multi-turn LLM conversation? Keeping full history gave far better recall than summarizing: in one evaluation, retaining the entire conversation let the model answer specific detail questions correctly 95% of the time, versus only 32% when the context was summarized first. Summarization strips out details the model later needs, causing it to miss facts even though it reduces the visible context size. _Developers weighing memory strategies for chat agents can track findings like this on daily.dev._ ### Does pure semantic search work for retrieving facts from very long documents? No, semantic (dense) search alone breaks down at scale: it retrieved facts well up to around 200k tokens with roughly 80% accuracy, but dropped to 0% recall once context reached 400k tokens with facts buried in the middle. Keyword search (BM25) held steady at 100% recall in the same test, which is why a hybrid dense-plus-BM25 approach performed best overall. _Teams designing retrieval pipelines for long documents can follow comparisons like this on daily.dev._ ## Similar posts on daily.dev - [How Compaction Works in Pi](https://daily.dev/posts/how-compaction-works-in-pi-e7tkghbxc) · Hacker News · 0 upvotes · 0 comments - [Context Management for DeepAgents](https://daily.dev/posts/context-management-for-deepagents-y0qa0czmt) · LangChain · 0 upvotes · 0 comments - [RAG vs. Long Context in 2026: When Retrieval Still Wins](https://daily.dev/posts/rag-vs-long-context-in-2026-when-retrieval-still-wins-ns3byqn4i) · Medium · 0 upvotes · 0 comments - [Context Window Management for Long-Running Agents: Strategies and Tradeoffs](https://daily.dev/posts/context-window-management-for-long-running-agents-strategies-and-tradeoffs-btfwkdimp) · Machine Learning Mastery · 0 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-inference](https://daily.dev/tags/ai-inference), [#context-engineering](https://daily.dev/tags/context-engineering), [#langchain](https://daily.dev/tags/langchain), [#rag](https://daily.dev/tags/rag) [View this post on daily.dev](https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Context Engineering in 2026 — Louis-François Bouchard, Omar Solano \u0026 Samridhi Vaid, Towards AI","url":"https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog"},"datePublished":"2026-08-17T16:45:42.146Z","dateModified":"2026-10-15T02:22:07.898Z","description":"A workshop from Towards AI walks through real production experiments on managing LLM context windows for an AI tutor product used by their students. The team...","image":"https://i.ytimg.com/vi/WP3hjUXd918/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/WP3hjUXd918/sddefault.jpg","isAccessibleForFree":true,"articleSection":"AI Engineer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"AI Engineer","logo":"https://media.daily.dev/image/upload/s--u5PucxNT--/f_auto/v1724338940/logos/aidotengineer","url":"https://daily.dev/sources/aidotengineer"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-inference,context-engineering,langchain,rag","timeRequired":"PT63M","video":{"@type":"VideoObject","name":"Context Engineering in 2026 — Louis-François Bouchard, Omar Solano \u0026 Samridhi Vaid, Towards AI","description":"A workshop from Towards AI walks through real production experiments on managing LLM context windows for an AI tutor product used by their students. The team...","thumbnailUrl":"https://i.ytimg.com/vi/WP3hjUXd918/sddefault.jpg","uploadDate":"2026-08-17T16:45:42.146Z","duration":"PT63M","url":"https://api.daily.dev/r/pSjpEiNoG","embedUrl":"https://www.youtube.com/embed/WP3hjUXd918"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"AI Engineer","item":"https://daily.dev/sources/aidotengineer"},{"@type":"ListItem","position":3,"name":"Context Engineering in 2026 — Louis-François Bouchard, Omar Solano \u0026 Samridhi Vaid, Towards AI"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/context-engineering-in-2026-louis-fran-ois-bouchard-omar-solano-samridhi-vaid-towards-ai-psjpeinog#faq","mainEntity":[{"@type":"Question","name":"Does compacting or summarizing chat history actually save money when using LLM prompt caching?","acceptedAnswer":{"@type":"Answer","text":"No, in most cases keeping the full chat history is cheaper than compacting. Summarizing or truncating context breaks the provider's prompt cache because it creates a new context, forcing the model to pay full price for those tokens again. Since cached tokens can cost up to 50 times less on some APIs like DeepSeek, compaction only pays off if it compresses the context by more than 50x without losing quality. Anyone tuning agent costs around prompt caching can find hands-on breakdowns like this through daily.dev."}},{"@type":"Question","name":"Does using RAG or full chat history give better recall than summarization in a multi-turn LLM conversation?","acceptedAnswer":{"@type":"Answer","text":"Keeping full history gave far better recall than summarizing: in one evaluation, retaining the entire conversation let the model answer specific detail questions correctly 95% of the time, versus only 32% when the context was summarized first. Summarization strips out details the model later needs, causing it to miss facts even though it reduces the visible context size. Developers weighing memory strategies for chat agents can track findings like this on daily.dev."}},{"@type":"Question","name":"Does pure semantic search work for retrieving facts from very long documents?","acceptedAnswer":{"@type":"Answer","text":"No, semantic (dense) search alone breaks down at scale: it retrieved facts well up to around 200k tokens with roughly 80% accuracy, but dropped to 0% recall once context reached 400k tokens with facts buried in the middle. Keyword search (BM25) held steady at 100% recall in the same test, which is why a hybrid dense-plus-BM25 approach performed best overall. Teams designing retrieval pipelines for long documents can follow comparisons like this on daily.dev."}}]} ```