mirror of
https://github.com/diegosouzapw/OmniRoute.git
synced 2026-10-19 13:30:37 +00:00
Validated in a combined board with the other PRs of this wave on release/v3.8.52: the PR's own tests pass; typecheck:core, open-sse typecheck, api-typecheck, complexity, changelog-integrity, env-doc-sync, test-discovery, vitest-exclusions, error-helper, test-masking, lockfile, deps, gate-manifest, cli-i18n and i18n coverage green. Any file-size ceiling or stryker registration this wave needs lands in the wave's follow-up PR. Thank you for the contribution!
This commit is contained in:
parent
d7070a5e96
commit
c69dce259f
1 changed files with 59 additions and 0 deletions
|
|
@ -911,6 +911,65 @@ Chain: production-fallback
|
|||
|
||||
---
|
||||
|
||||
### Common Provider Combinations & Routing Patterns
|
||||
|
||||
Here are example patterns for combining multiple providers and routing between them in OmniRoute:
|
||||
|
||||
#### 1. Coding Agent Combo: High-End Reasoning with Cost/Speed Fallback
|
||||
|
||||
Ideal for coding agents (OpenCode, Claude Code, Cursor, Cline). Routes initially to frontier reasoning models, falling back to fast coding models on quota exhaustion or errors.
|
||||
|
||||
- **Dashboard**: Combos → New Combo → Name: `agent-coding` → Strategy: `Priority`
|
||||
- **Models**:
|
||||
1. `claude/claude-sonnet-4-6` (Primary coding agent)
|
||||
2. `openai/gpt-4o` (Secondary high-capacity fallback)
|
||||
3. `deepseek/deepseek-v4-flash` (High-efficiency, cost-effective fallback)
|
||||
|
||||
```bash
|
||||
# Example via API
|
||||
curl -X POST http://localhost:20128/api/combos \
|
||||
-H "Authorization: Bearer $OMNIROUTE_MANAGE_KEY" \
|
||||
-H "Content-Type: application/json" \
|
||||
-d '{
|
||||
"name": "agent-coding",
|
||||
"strategy": "priority",
|
||||
"models": [
|
||||
{ "model": "claude/claude-sonnet-4-6" },
|
||||
{ "model": "openai/gpt-4o" },
|
||||
{ "model": "deepseek/deepseek-v4-flash" }
|
||||
]
|
||||
}'
|
||||
```
|
||||
|
||||
#### 2. Free-Tier Auto Failover Combo
|
||||
|
||||
Chains multiple free-tier and keyless providers to maximize uptime without API costs.
|
||||
|
||||
- **Strategy**: `Least Used` or `Round Robin` (distributes load across quotas)
|
||||
- **Models**:
|
||||
1. `groq/llama-3.3-70b-versatile`
|
||||
2. `gemini/gemini-2.5-flash`
|
||||
3. `cerebras/gpt-oss-120b`
|
||||
|
||||
```bash
|
||||
# Example via CLI
|
||||
omniroute combo create free-stack \
|
||||
--models "groq/llama-3.3-70b-versatile,gemini/gemini-2.5-flash,cerebras/gpt-oss-120b" \
|
||||
--strategy round-robin
|
||||
```
|
||||
|
||||
#### 3. Multimodal / Vision & Text Pipeline
|
||||
|
||||
Pairing specialized vision models with high-speed text generation for workflows involving image understanding and code generation.
|
||||
|
||||
- **Pattern**: A `Priority` combo that lists vision-capable models first and a high-throughput text/code model last.
|
||||
- **Models**:
|
||||
1. `gemini/gemini-2.5-pro` (Strong image/multimodal comprehension)
|
||||
2. `openai/gpt-4o` (Balanced vision and tool usage)
|
||||
3. `deepseek/deepseek-v4-flash` (Text/code generation)
|
||||
|
||||
---
|
||||
|
||||
### Resilience & Circuit Breakers
|
||||
|
||||
Configure via **Dashboard → Settings → Resilience**.
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue