pawan/config/
routing.rs

1use serde::{Deserialize, Serialize};
2
3use super::provider::LlmProvider;
4
5/// Task-type model routing — use different models for different task categories.
6///
7/// # Example (pawan.toml)
8/// ```toml
9/// [models]
10/// code = "qwen/qwen3.5-122b-a10b"                  # best for code generation
11/// orchestrate = "minimaxai/minimax-m2.5"            # best for tool calling
12/// execute = "mlx-community/Qwen3.5-9B-OptiQ-4bit"  # fast local execution
13/// ```
14#[derive(Debug, Clone, Default, Serialize, Deserialize)]
15pub struct ModelRouting {
16    /// Model for code generation tasks (implement, refactor, write tests)
17    pub code: Option<String>,
18    /// Model for orchestration tasks (multi-step tool chains, analysis)
19    pub orchestrate: Option<String>,
20    /// Model for simple execution tasks (bash, write_file, cargo test)
21    pub execute: Option<String>,
22}
23
24impl ModelRouting {
25    /// Select the best model for a given task based on keyword analysis.
26    /// Returns None if no routing matches (use default model).
27    pub fn route(&self, query: &str) -> Option<&str> {
28        let q = query.to_lowercase();
29
30        // Code generation patterns
31        if self.code.is_some() {
32            let code_signals = [
33                "implement",
34                "write",
35                "create",
36                "refactor",
37                "fix",
38                "add test",
39                "add function",
40                "struct",
41                "enum",
42                "trait",
43                "algorithm",
44                "data structure",
45            ];
46            if code_signals.iter().any(|s| q.contains(s)) {
47                return self.code.as_deref();
48            }
49        }
50
51        // Orchestration patterns
52        if self.orchestrate.is_some() {
53            let orch_signals = [
54                "search", "find", "analyze", "review", "explain", "compare", "list", "check",
55                "verify", "diagnose", "audit",
56            ];
57            if orch_signals.iter().any(|s| q.contains(s)) {
58                return self.orchestrate.as_deref();
59            }
60        }
61
62        // Execution patterns
63        if self.execute.is_some() {
64            let exec_signals = [
65                "run", "execute", "bash", "cargo", "test", "build", "deploy", "install", "commit",
66            ];
67            if exec_signals.iter().any(|s| q.contains(s)) {
68                return self.execute.as_deref();
69            }
70        }
71
72        None
73    }
74}
75
76/// Cloud fallback configuration for hybrid local+cloud model routing.
77///
78/// When the primary provider (typically a local model via OpenAI-compatible API)
79/// fails or is unavailable, pawan automatically falls back to this cloud provider.
80/// This enables zero-cost local inference with cloud reliability as a safety net.
81///
82/// # Example (pawan.toml)
83/// ```toml
84/// provider = "openai"
85/// model = "Qwen3.5-9B-Q4_K_M"
86///
87/// [cloud]
88/// provider = "nvidia"
89/// model = "mistralai/devstral-2-123b-instruct-2512"
90/// ```
91#[derive(Debug, Clone, Serialize, Deserialize)]
92pub struct CloudConfig {
93    /// Cloud LLM provider to fall back to (nvidia or openai)
94    pub provider: LlmProvider,
95    /// Primary cloud model to try first on fallback
96    pub model: String,
97    /// Additional cloud models to try if the primary cloud model also fails
98    #[serde(default)]
99    pub fallback_models: Vec<String>,
100}
pawan/config/routing.rs

pawan/config/
routing.rs
