Latest / Big Iron Radio / Homelab Brief - Semantic Camera Search with Frigate and Ollama

Transcript

Homelab Brief - Semantic Camera Search with Frigate and Ollama · Big Iron Radio

• 0:02
  Welcome to the Homelab Brief from Big Iron.
• 0:06
  Today, adding semantic search and AI descriptions to your frigate cameras kept fully local with Alama.
• 0:14
  Quick note, this is an AI voice drawn from a Big Iron guide.
• 0:18
  Links below.
• 0:20
  First, untangle two features people lump together.
• 0:23
  Semantic search and gen AI descriptions are separate, and they have almost nothing in common resource-wise.
• 0:29
  The search embeddings run on CPU, or iGPU for Penis.
• 0:35
  The gen AI captions need real GPU time through Alama.
• 0:39
  Budget them separately.
• 0:41
  The local win is one line.
• 0:44
  Set provider Alama in the genite block, and your captions stay fully local.
• 0:49
  No open AI or Gemini key required, and this is the part that matters for cameras no footage metadata ever leaves your network.
• 0:57
  On hardware, 8GB of RAM is the realistic floor for decent captions.
• 1:04
  In that range, Minikpum V8B and QN2VL7B are the stronger picks.
• 1:10
  On a tighter budget, Moondream 1, 8B and Lavify 3 work with weaker output.
• 1:15
  Match the model to the card.
• 1:18
  The gotcha that bites people.
• 1:20
  Scope gen AI by camera, and by required zones, before you enable it globally.
• 1:24
  Turn it on unscoped across a busy camera set, and you create a GPU queue backlog captions, pile up and arrive long after they're useful.
• 1:34
  Describe the front door, not every leaf blowing past every camera.
• 1:39
  And one setup check.
• 1:41
  Make sure the Alama base URL is actually reachable from inside the frigate container, not just from your desktop, that's the number one, it silently does nothing, cause.
• 1:51
  The full walkthroughs in the show notes.
• 1:54
  That's your brief.