WEBVTT

1
00:00:00.000 --> 00:00:05.890
Let's take a closer look at how SemanticRouter.NET decides where each prompt goes, and what happens when something goes wrong.

2
00:00:05.844 --> 00:00:08.193
Every request comes down to two questions.

3
00:00:08.193 --> 00:00:08.529
First:

4
00:00:08.529 --> 00:00:10.543
what does this prompt actually need?

5
00:00:10.543 --> 00:00:13.116
That's a tier, plus things like code or tools.

6
00:00:13.116 --> 00:00:13.508
Second:

7
00:00:13.508 --> 00:00:16.864
which of your models is the cheapest one that can handle it?

8
00:00:20.423 --> 00:00:22.269
The first answer comes from routes.

9
00:00:22.269 --> 00:00:26.859
A route is just a handful of example prompts, plus the tier that kind of request needs.

10
00:00:26.859 --> 00:00:32.503
You start with eight of them, from small talk at the bottom, up to debugging and agentic coding at the top.

11
00:00:33.454 --> 00:00:37.042
When a prompt comes in, the router compares it with every example.

12
00:00:37.042 --> 00:00:42.044
This one is closest to complex code, so it needs tier four, and a model that can write code.

13
00:00:42.299 --> 00:00:43.518
Now something smaller:

14
00:00:43.518 --> 00:00:45.789
a regular expression for email addresses.

15
00:00:45.789 --> 00:00:49.889
That's simple code, so tier two is plenty, and the local coder goes first.

16
00:00:50.289 --> 00:00:55.728
Ask for the same thing, but add a list of things to reason through, and the complexity score goes up.

17
00:00:55.728 --> 00:01:00.036
Long prompts, code, structure, and phrases like step by step all push it higher.

18
00:01:00.036 --> 00:01:02.459
Past a threshold, the prompt moves up a tier.

19
00:01:06.702 --> 00:01:08.482
Your models sit on tiers too.

20
00:01:08.482 --> 00:01:13.086
This page shows which routes need each tier, and which models can serve it.

21
00:01:13.086 --> 00:01:16.892
If a tier has nothing that can serve it, you'll see it in red.

22
00:01:18.079 --> 00:01:20.888
Then a strategy ranks the models that qualify.

23
00:01:20.888 --> 00:01:22.842
The default is cheapest capable.

24
00:01:22.842 --> 00:01:29.742
Local models go first, and the most capable one that's already loaded wins, so nothing waits for a model to load.

25
00:01:29.742 --> 00:01:33.284
Then your subscriptions, lowest tier first, to save quota.

26
00:01:33.284 --> 00:01:34.505
Paid APIs come last.

27
00:01:34.505 --> 00:01:39.329
Switch to auto best, and the same prompt goes straight to your strongest model.

28
00:01:43.370 --> 00:01:45.591
Coding agents send a lot more than prompts.

29
00:01:45.591 --> 00:01:50.550
Every Claude Code turn carries its tools, so only models that can call tools are in the running.

30
00:01:50.551 --> 00:01:53.316
Here, the local model asks Claude Code to open a file.

31
00:01:53.316 --> 00:01:56.031
Claude Code runs the tool, and sends the result back.

32
00:01:56.031 --> 00:02:01.971
And the router hands that result to the same model that asked for it, so a turn never gets split between two models.

33
00:02:01.972 --> 00:02:03.734
And you can always take over.

34
00:02:03.734 --> 00:02:09.142
Start a message with at opus, at local, or at tier two, and the router does exactly that.

35
00:02:12.477 --> 00:02:14.527
Some data should never leave your machine.

36
00:02:14.527 --> 00:02:18.823
Turn on secret detection, and once a conversation contains a credential, it stays local.

37
00:02:18.823 --> 00:02:20.287
Even if you've asked for Opus.

38
00:02:26.230 --> 00:02:31.456
When a model goes down, or runs out of quota, it cools down, and the next one picks up the work.

39
00:02:31.456 --> 00:02:34.940
The endpoints table shows what's cooling down, and for how long.

40
00:02:39.387 --> 00:02:41.252
Every decision explains itself.

41
00:02:41.252 --> 00:02:46.307
The activity log shows the route, the tier, and every model the request was sent to.

42
00:02:47.495 --> 00:02:50.964
And the evaluate page runs a set of labelled prompts through your configuration.

43
00:02:50.964 --> 00:02:55.735
You'll see how often the router gets it right, and exactly where it gets it wrong, before you change anything.

44
00:02:56.595 --> 00:02:58.883
So that's how a prompt gets routed.

45
00:02:58.883 --> 00:03:01.825
Next time, we'll tune it for your own models.

46
00:03:01.825 --> 00:03:05.225
SemanticRouter.NET is open source, and MIT licensed.