Routing between combos

[RodriAndreotti]

Hi, there!

First of all, great project!
Congrats!

I've installed omniroute yesterday and have created 4 combos with differente purposes models.

How could I implement dynamic routing between these four combinations based on the prompt?

Something like an AI classifier.

Is there anything like this available today?

I’ve seen that it’s possible to create routes based on glob patterns, but I’m not sure if that would lead to too many false positives for each model.

My idea would be something like:

Architecture model
Execution model
Low-cost model for batch processing, and
High-end model (Opus or similar) for conflict resolution among the others

Does anyone have suggestions for these kinds of scenarios?

[diegosouzapw]

Hey @RodriAndreotti!

Great question, and congrats on getting set up with 4 purpose-specific combos - that is exactly the kind of advanced setup OmniRoute is built for!

OmniRoute already has features for dynamic prompt-based routing:

1. Model-Combo Mappings (Wildcard Router)
In Dashboard -> Combos, you can set up model-combo mappings that route requests based on the model name in the request. For example:

• gpt-4* -> your high-end combo
• • gpt-3.5* -> your low-cost batch combo
• • claude* -> your architecture combo
    2. Intent Classifier and Task-Aware Router
    OmniRoute has a built-in intent classifier and task-aware router that can analyze the prompt content and route to appropriate combos. These work with the combo routing engine automatically.

3. Combo Strategies
Each combo supports 13 different strategies. For your use case, you might want:

• priority for your high-end resolver (Opus)
• • cost-optimized for batch processing
• • context-optimized for architecture decisions
    4. MCP Tools
    If you use OmniRoute's MCP server, the best_combo_for_task and simulate_route tools let you programmatically test routing decisions before committing.

The simplest approach for your 4-combo setup would be to use model-combo mappings with specific model aliases. Let me know if you would like help configuring this!

[RodriAndreotti]

Hey!
Great news so... :-)

As workaround I've configured an classifier on kilo code pointing to a free open model, this model classifies my prompt and then redirect to correct combo, but if i can do it without dependency of local settings will be better.

If you can provide a tutorial to make this config nativelly.

And if i can help in some way with de project let me know

[diegosouzapw]

Hey @RodriAndreotti! Update on this: there's a path forward without needing your external classifier.

OmniRoute actually already ships the building blocks -- two sync, sub-ms, keyword-based classifiers:

• open-sse/services/intentClassifier.ts -- detects code | math | reasoning | creative | simple | medium from prompts in 9 languages (EN, PT-BR, ES, ZH, JA, RU, DE, KO, AR).
• open-sse/services/taskAwareRouter.ts -- detects 7 task types (coding / creative / analysis / vision / summarization / background / chat) and maps each to a preferred model.

The gap: taskAwareRouter today maps each task type to a single model, not to a combo. So your scenario (4 purpose-specific combos -> auto-pick by prompt) isn't directly wired yet -- but the detector is already there and runs in <1ms.

I've opened #2834 to track the extension -- add combo aliases as valid targets in taskModelMap and surface a "Smart Router" tab in /dashboard/combos where you'd map coding -> your-coding-combo, creative -> your-creative-combo, etc.

Until that ships, the workaround you set up (small free model classifying client-side) is the right shape. Will ping you here when the inter-combo router lands.

[RodriAndreotti]

Hey, @diegosouzapw!
Looks great, i'll try it.
And in my agent i'll need to use omnirout/auto or something like this?

[diegosouzapw]

Yep @RodriAndreotti -- once the Smart Router lands (#2834), the agent side stays simple: just point your client at a single stable model name (e.g. omniroute-smart or gpt-5 -- whatever you alias) and OmniRoute decides which combo to enter based on the prompt content.

Two ways to consume it:

1. Auto-Combo aware alias -- in /dashboard/models -> New Alias, create omniroute-smart and tick "Use Smart Router". Your agent sends requests to model omniroute-smart and the classifier picks coding-combo / creative-combo / etc per request.
2. Catch-all model-name mapping -- map gpt-* (glob) -> Smart Router. Then any model your agent already requests routes through the classifier first.

For now (until #2834 ships), the cleanest pre-existing path is the alias-per-role pattern I described in #540's reply -- one model alias per agent role, each backed by a single combo. It's not prompt-classified but it gets you per-agent fallback chains today without the external classifier dependency.

I'll ping back here when the Smart Router PR lands.

[RodriAndreotti]

Thanks a lot!
And, again, great work!

[diegosouzapw]

Cheers @RodriAndreotti -- thank you! I'll ping you right here the moment the Smart Router (#2834) lands so you can drop the external classifier. Appreciate the kind words and the well-scoped use case.

[diegosouzapw]

As promised, @RodriAndreotti -- pinging you here because the Smart Router (#2834) has landed, so you can drop your external Kilo classifier.

It shipped as task-aware routing that now accepts combos as targets (not just single models). The built-in keyword classifier (intentClassifier / taskAwareRouter, 9 languages) detects the task type per prompt and overrides the request to the combo you mapped.

How to wire it to your 4 combos:

1. Enable it and map each task type to one of your combos via PUT /api/settings/task-routing:

   {
     "enabled": true,
     "taskModelMap": {
       "coding":   "combo/execution",
       "analysis": "combo/architecture",
       "background":"combo/low-cost-batch",
       "reasoning":"combo/high-end"
     }
   }

   (or via the dashboard Settings -> Task Routing panel).
2. Your agent keeps sending normal requests -- when task routing is enabled, every request is classified and, if the detected task type has a mapping, OmniRoute overrides to that combo automatically (src/sse/handlers/chat.ts:354). No per-request alias needed.

Note: this corrects my earlier "New Alias -> Use Smart Router" sketch -- the shipped mechanism is the global Task Routing toggle + taskModelMap, which is cleaner for your 4-combo setup. Give it a spin and let me know how the classification holds up on your prompts.