--- title: "Your RAG is lying to you" newsletter: "MLOps Community" date: 2026-04-02 source: https://aaif.live/newsletters/mlopscommunity/2026-04-02-your-rag-is-lying-to-you --- # Your RAG is lying to you *Plus: the shift that makes developers obsolete (and what replaces them)* *MLOps Community — Agentic AI Foundation, 2026-04-02* Could this finally make Slack search usable [https://techcrunch.com/2026/03/31/salesforce-announces-an-ai-heavy-makeover-for-slack-with-30-new-features/]? ## Retirement Party for One Single-agent systems are already legacy - coordination is the real product now. What are you building: assistants or systems? [ASSISTANTS](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=Single-agent+systems+are+already+legacy+-+coordination+is+the+real+product+now.What+are+you+building%3A+assistants+or+systems%3F&y=ASSISTANTS&n=SYSTEMS&oneclick=true&accept=true) [SYSTEMS](https://gatewaze.mlops.community/offer/surveys/?sid=yesno&question=Single-agent+systems+are+already+legacy+-+coordination+is+the+real+product+now.What+are+you+building%3A+assistants+or+systems%3F&y=ASSISTANTS&n=SYSTEMS&oneclick=true&accept=true) ## Harness the solution Only 8% think a better model solves it. The other 92% know the harness is where things break. ## OpenXData 2026 - Virtual Event on Open Data Architectures OpenXData is back. April 29. 9,000 showed up last year. This year we're going bigger. Demetrios Brinkman is hosting (shenanigans guaranteed), with a speaker lineup pulled straight from Uber, Meta, IBM, Microsoft, Alibaba, Booking, Anthropic, Cloudera, JD.com [http://JD.com]. Not the people who blog about these systems. The people who built them. Two tracks. All signal, no filler - Presto, Velox, Hudi, Polaris, Gluten, and where data infrastructure goes in a world that's suddenly AI-first. On the agenda: * AI-native data platforms - the actual implementation, not the pitch deck * Data engineering for AI workloads, in production * Cost and performance at scale, where things get genuinely hard * Open data stacks that actually work together Panels, workshops, deep dives, all designed for people shipping real systems. No intro talks. No content you've seen before. Just what's working right now, from the teams running it live. If you're building in production - this one's for you. REGISTER TODAY [REGISTER TODAY](https://www.openxdata.ai/) [https://www.openxdata.ai/](https://www.openxdata.ai/) ## Curated finds to help you stay ahead ## arrowspace: Vector Spaces and Graph Wiring RAG can look sharp at the top of the ranking and still quietly trap an agent in the same few documents. This discussion looks at a different fix: changing the search itself. * Standard vector search is strong on the first few hits, then falls off fast. The argument here is that feature relationships hold extra structure that plain geometric similarity misses. * A graph-based search layer can recover some of that lost signal, making smaller embeddings behave more like much larger ones and giving agents more useful paths through relevant data. * That same structure may help with ranking tails, long-term context, and dataset evaluation across MLOps workflows. The result is a search stack that gives reasoning systems more room to move without drifting into junk. [https://podcasts.apple.com/gb/podcast/arrowspace-vector-spaces-and-graph-wiring/id1505372978?i=1000757765927](https://podcasts.apple.com/gb/podcast/arrowspace-vector-spaces-and-graph-wiring/id1505372978?i=1000757765927) [https://home.mlops.community/home/videos/arrowspace-vector-spaces-and-graph-wiring](https://home.mlops.community/home/videos/arrowspace-vector-spaces-and-graph-wiring) [https://open.spotify.com/episode/6cBeHFP1JhfAbOz6zLQnYV?si=0ceea6266fd243ef](https://open.spotify.com/episode/6cBeHFP1JhfAbOz6zLQnYV?si=0ceea6266fd243ef) ## This One Shift Makes Developers Obsolete Some reflections with an attendee from our recent South Bay event on coding agents - and how quickly things are shifting. * The move is from single assistants to coordinated systems: parallel agents, review agents, hooks, and eval layers that manage work more like infrastructure. * Results depend less on model choice and more on the harness around it - specs, skills, monitoring, permissions, and targeted human review. * That also changes ownership. Faster output is useful, but someone still has to know what to trust, what to inspect, and where the workflow can fail. The focus is shifting from writing code faster to managing how that code gets produced and validated. If this got you thinking, our next event is April 14 in Seattle - early bird pricing closes this Friday [https://luma.com/ai-agents-summit-seattle]. [https://podcasts.apple.com/gb/podcast/spec-driven-development-workflows-and-the-recent/id1505372978?i=1000758521396](https://podcasts.apple.com/gb/podcast/spec-driven-development-workflows-and-the-recent/id1505372978?i=1000758521396) [https://home.mlops.community/home/videos/this-one-shift-makes-developers-obsolete](https://home.mlops.community/home/videos/this-one-shift-makes-developers-obsolete) [https://open.spotify.com/episode/4psJMG6Y3WTNoUehjzqdZO?si=fbc13facb9364b58](https://open.spotify.com/episode/4psJMG6Y3WTNoUehjzqdZO?si=fbc13facb9364b58) ## Operationalizing AI Agents: From Experimentation to Production // Databricks Roundtable A lot of agent demos look impressive right up until they touch real users, real data, and real workflows. This panel gets into what happens after that. * Teams are finding that broad, do-everything agents rarely hold up. Narrower scopes, clearer product thinking, and tighter task design tend to survive contact with production better. * Trust comes from layers, not hope: evals, traces, integration tests, human checks, and limited permissions all matter once agents start making decisions inside live systems. * Internal agents can still be hugely useful, especially for read-heavy tasks, but most teams are keeping write access constrained and treating reliability as an ongoing process. The pattern is less about handing work to one smart agent and more about building enough structure around it that failure stays manageable. [https://podcasts.apple.com/gb/podcast/operationalizing-ai-agents-from-experimentation-to/id1505372978?i=1000758267017](https://podcasts.apple.com/gb/podcast/operationalizing-ai-agents-from-experimentation-to/id1505372978?i=1000758267017) [https://home.mlops.community/home/videos/operationalizing-ai-agents-from-experimentation-to-production-databricks-roundtable](https://home.mlops.community/home/videos/operationalizing-ai-agents-from-experimentation-to-production-databricks-roundtable) [https://open.spotify.com/episode/5vHu7BfPfFVpuXofV9bv9k?si=0388939ea5864b46](https://open.spotify.com/episode/5vHu7BfPfFVpuXofV9bv9k?si=0388939ea5864b46) ## Engineering the Memory Layer For An AI Agent To Navigate Large-scale Event Data Event data usually ends up scattered across folders, pages, and formats, which makes anything beyond basic search painful. This piece shows how to structure that mess so an agent can query it properly. * The focus is on the memory layer. A graph-based schema encodes relationships upfront, so the agent can traverse Speaker → Talk → Topic instead of guessing via vector search alone. * The pipeline combines metadata, transcript chunks, and embeddings into connected descriptor sets, enabling queries that mix filters, joins, and semantic search in a single step. * That structure supports more precise retrieval, including constrained search within subsets of talks and linking results to exact moments in videos. The key takeaway is that agent performance is heavily shaped by how the data is organised, not just the model or orchestration layer. [Read the blog](https://home.mlops.community/home/blogs/engineering-the-memory-layer-for-an-ai-agent-to-navigate-large-scale-event-data) ## IN PERSON EVENTS * New York [https://luma.com/uk9vvgal] - April 8 * Seattle [https://luma.com/ai-agents-summit-seattle] - April 14 * Amsterdam [https://luma.com/i7o74wuo] - April 21 * San Francisco [https://luma.com/p8d635rz] - May 15 ## VIRTUAL EVENTS * Coding Agents Lunch and Learn [https://home.mlops.community/home/events/coding-agents-lunch-and-learn-session-7-javid0cpse] - April 3 ## TODO: delete before publish [https://forms.gle/8EDvXGizxyFVKfwy8](https://forms.gle/8EDvXGizxyFVKfwy8) --- Source: https://aaif.live/newsletters/mlopscommunity/2026-04-02-your-rag-is-lying-to-you