--- title: Semantic Blindness: 500,000 Sensors Confused an LLM -... description: Engineers at Phaidra (Fedra) describe a problem they call 'semantic blindness': LLMs fail when given 500,000+ sensor and equipment names from AI factory data... canonical: https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra | daily.dev og:description: Engineers at Phaidra (Fedra) describe a problem they call 'semantic blindness': LLMs fail when given 500,000+ sensor and equipment names from AI factory data... og:url: https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw og:image: https://api.daily.dev/og/posts/e7NwAiLQW.png og:image:alt: Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra 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. # Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra **[AI Engineer](https://daily.dev/sources/aidotengineer)** · 16 min read · 1 upvotes · 0 comments ## Summary Engineers at Phaidra (Fedra) describe a problem they call 'semantic blindness': LLMs fail when given 500,000+ sensor and equipment names from AI factory data centers. Naive approaches like stuffing names into context windows or using vector RAG both break at scale due to context limits, frequency penalties, and near-identical names. Their solution uses a hierarchical tree structure of the physical infrastructure, separating LLM planning from deterministic retrieval. The LLM generates structured search patterns and scope filters; a deterministic resolver executes set operations for perfect recall. This reduces token cost from 116 million to ~9,000 per query regardless of system size. The key insight: LLMs excel at parsing ambiguous intent and planning, but structured data traversal, exact matching, and set logic should be deterministic code. They frame this as 'AI-native software starting at Software 3.0 and maturing toward 1.0' as it scales to production. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why does an LLM fail to correctly find equipment when given hundreds of thousands of near-identical sensor names in a data center? Correctness degrades because context windows saturate, near-identical names (like Chiller 6 vs Chiller 7) cause poor vector search recall, and LLMs hit internal frequency penalties that suppress output when repeatedly listing similar tokens. Testing showed correctness dropping from 80% at 64 GPUs to about 30% at 460,000 GPUs using naive LLM or RAG-based approaches, with sharded parallel calls also producing hallucinated phantom equipment and silently dropped real equipment. _Anyone debugging LLM recall failures on large structured datasets can find similar architecture writeups curated on daily.dev._ ### How can I make an LLM query system scale to hundreds of thousands of pieces of equipment without ballooning token costs? Exploit the hierarchical tree structure of the data (data center, hall, aisle, row, rack, GPU) so the LLM only reasons over a small, constant-size summary of tree paths rather than every individual name. The LLM outputs a structured plan naming a scope and filter, and a deterministic backend resolves it using pre-indexed subtrees and set operations, keeping query cost flat at roughly 9,000 tokens whether the system has 64 GPUs or 460,000 GPUs. _Engineers designing LLM systems for large hierarchical infrastructure can track architecture patterns like this via daily.dev._ ### What is Karpathy's software 1.0 versus 3.0 framing and how does it apply to production LLM system design? Software 1.0 is deterministic, exact code you write, while software 3.0 is behavior prompted out of an LLM, which is flexible but fuzzy. Legacy software tends to drift from 1.0 toward 3.0, but building new AI-native systems for large, well-structured data should invert that: start with everything in the LLM to validate the idea quickly, then move anything with clear structure, exact retrieval, counting, or deduplication into deterministic 1.0 code, keeping only ambiguous judgment calls in the model. _Teams deciding how much logic to hand an LLM versus deterministic code can follow related engineering debates on daily.dev._ ## Similar posts on daily.dev - [Semantic Search for AI Agents at Scale: Retrieval and Ranking for LinkedIn’s Hiring Assistant](https://daily.dev/posts/semantic-search-for-ai-agents-at-scale-retrieval-and-ranking-for-linkedin-s-hiring-assistant-ltbv77qsk) · LinkedIn Engineering · 0 upvotes · 0 comments - [Ontology driven Dimensional Modeling](https://daily.dev/posts/ontology-driven-dimensional-modeling-jkjpvwzta) · dltHub · 1 upvotes · 0 comments - [Researchers discover a shortcoming that makes LLMs less reliable](https://daily.dev/posts/researchers-discover-a-shortcoming-that-makes-llms-less-reliable-wiiiapmcn) · MIT News · 3 upvotes · 0 comments - [What If We Don't Need the Semantic Layer?](https://daily.dev/posts/what-if-we-don-t-need-the-semantic-layer--oc091fmdh) · MotherDuck · 5 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#vector-search](https://daily.dev/tags/vector-search) [View this post on daily.dev](https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw) ```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":"Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra","url":"https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw"},"datePublished":"2026-07-12T16:06:32.828Z","dateModified":"2026-09-14T05:59:54.766Z","description":"Engineers at Phaidra (Fedra) describe a problem they call 'semantic blindness': LLMs fail when given 500,000+ sensor and equipment names from AI factory data...","image":"https://i.ytimg.com/vi/EUsPvBeIx70/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/EUsPvBeIx70/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/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,vector-search","timeRequired":"PT16M","video":{"@type":"VideoObject","name":"Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra","description":"Engineers at Phaidra (Fedra) describe a problem they call 'semantic blindness': LLMs fail when given 500,000+ sensor and equipment names from AI factory data...","thumbnailUrl":"https://i.ytimg.com/vi/EUsPvBeIx70/sddefault.jpg","uploadDate":"2026-07-12T16:06:32.828Z","duration":"PT16M","url":"https://api.daily.dev/r/e7NwAiLQW","embedUrl":"https://www.youtube.com/embed/EUsPvBeIx70"}} {"@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":"Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/semantic-blindness-500-000-sensors-confused-an-llm---raahul-singh-van-levstik-phaidra-e7nwailqw#faq","mainEntity":[{"@type":"Question","name":"Why does an LLM fail to correctly find equipment when given hundreds of thousands of near-identical sensor names in a data center?","acceptedAnswer":{"@type":"Answer","text":"Correctness degrades because context windows saturate, near-identical names (like Chiller 6 vs Chiller 7) cause poor vector search recall, and LLMs hit internal frequency penalties that suppress output when repeatedly listing similar tokens. Testing showed correctness dropping from 80% at 64 GPUs to about 30% at 460,000 GPUs using naive LLM or RAG-based approaches, with sharded parallel calls also producing hallucinated phantom equipment and silently dropped real equipment. Anyone debugging LLM recall failures on large structured datasets can find similar architecture writeups curated on daily.dev."}},{"@type":"Question","name":"How can I make an LLM query system scale to hundreds of thousands of pieces of equipment without ballooning token costs?","acceptedAnswer":{"@type":"Answer","text":"Exploit the hierarchical tree structure of the data (data center, hall, aisle, row, rack, GPU) so the LLM only reasons over a small, constant-size summary of tree paths rather than every individual name. The LLM outputs a structured plan naming a scope and filter, and a deterministic backend resolves it using pre-indexed subtrees and set operations, keeping query cost flat at roughly 9,000 tokens whether the system has 64 GPUs or 460,000 GPUs. Engineers designing LLM systems for large hierarchical infrastructure can track architecture patterns like this via daily.dev."}},{"@type":"Question","name":"What is Karpathy's software 1.0 versus 3.0 framing and how does it apply to production LLM system design?","acceptedAnswer":{"@type":"Answer","text":"Software 1.0 is deterministic, exact code you write, while software 3.0 is behavior prompted out of an LLM, which is flexible but fuzzy. Legacy software tends to drift from 1.0 toward 3.0, but building new AI-native systems for large, well-structured data should invert that: start with everything in the LLM to validate the idea quickly, then move anything with clear structure, exact retrieval, counting, or deduplication into deterministic 1.0 code, keeping only ambiguous judgment calls in the model. Teams deciding how much logic to hand an LLM versus deterministic code can follow related engineering debates on daily.dev."}}]} ```