# Semantic Blindness: 500,000 Sensors Confused an LLM - Raahul Singh & Vanč Levstik, Phaidra

Source: [YouTube](https://www.youtube.com/watch?v=EUsPvBeIx70)  
Feed7 permalink: https://feed7.dev/p/semantic-blindness-500-000-sensors-confused-an-llm-raahul-singh-vanc-lev-159c4yr  
Published: Unknown  
Trust: Source Linked (source_linked)

## Why Included

Use LLMs to turn ambiguous requests into search plans, then resolve entities with deterministic indexes and set operations. This avoids context bloat and silent misses at production scale.

## Source Summary

Phaidra’s direct LLM approach fell from **80% correctness at 64 GPUs** to about **30% at 460,000 GPUs**. Its replacement uses a planner model, pre-indexed hierarchy and deterministic set operations, maintaining **100% correctness** in the reported tests.

## Practical Implication

Keep model judgment at the fuzzy boundary: interpret intent, choose scope and formulate filters. Move exact retrieval, counting, deduplication and intersections into code whenever your domain has a hierarchy, graph or schema.

## Agent-Ready Context

Phaidra’s direct LLM approach fell from **80% correctness at 64 GPUs** to about **30% at 460,000 GPUs**. Its replacement uses a planner model, pre-indexed hierarchy and deterministic set operations, maintaining **100% correctness** in the reported tests.

Keep model judgment at the fuzzy boundary: interpret intent, choose scope and formulate filters. Move exact retrieval, counting, deduplication and intersections into code whenever your domain has a hierarchy, graph or schema.

The results come from Phaidra’s workload: six production systems plus scaled tests involving data-center equipment. The claim of perfect recall depends on accurate indexes, filters and structured plans; vague requests may still require model-generated search patterns.

## Context Map

- Layer: agent
- Domains: data
- Topics: harness-engineering, tool-use, agent-reliability

## Uncertainty

- The results come from Phaidra’s workload: six production systems plus scaled tests involving data-center equipment. The claim of perfect recall depends on accurate indexes, filters and structured plans; vague requests may still require model-generated search patterns.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.