noether_engine/
providers.rs

1//! Provider factory — builds embedding and LLM providers from environment config.
2//!
3//! ## Provider priority (LLM and embeddings)
4//!
5//! 1. **Mistral native** (`MISTRAL_API_KEY` set) — calls `api.mistral.ai` directly.
6//!    Preferred for the European deployment stack: no Google Cloud dependency,
7//!    EU-hosted, GDPR-compliant. Set `MISTRAL_MODEL` to override the model.
8//!
9//! 2. **OpenAI** (`OPENAI_API_KEY` set) — calls OpenAI or any compatible API.
10//!    Override base URL with `OPENAI_API_BASE` for Ollama, Together, etc.
11//!
12//! 3. **Anthropic** (`ANTHROPIC_API_KEY` set) — calls `api.anthropic.com`.
13//!    LLM only (no embeddings).
14//!
15//! 4. **Vertex AI** (`VERTEX_AI_PROJECT` + credentials set) — Google Cloud.
16//!    Supports Mistral, Gemini, and Claude via Vertex Model Garden.
17//!
18//! 5. **Mock** (fallback) — deterministic hash-based embeddings, echo LLM.
19//!    Used in tests and when no cloud credentials are present.
20//!
21//! ## Environment variables
22//!
23//! | Variable | Description | Default |
24//! |---|---|---|
25//! | `MISTRAL_API_KEY` | Native Mistral API key (console.mistral.ai) | — |
26//! | `MISTRAL_MODEL` | Mistral model name | `mistral-small-latest` |
27//! | `MISTRAL_EMBEDDING_MODEL` | Mistral embedding model | `mistral-embed` |
28//! | `OPENAI_API_KEY` | OpenAI API key (or compatible) | — |
29//! | `OPENAI_MODEL` | OpenAI model name | `gpt-4o-mini` |
30//! | `OPENAI_API_BASE` | OpenAI-compatible base URL | `https://api.openai.com/v1` |
31//! | `OPENAI_EMBEDDING_MODEL` | OpenAI embedding model | `text-embedding-3-small` |
32//! | `ANTHROPIC_API_KEY` | Anthropic API key | — |
33//! | `ANTHROPIC_MODEL` | Anthropic model name | `claude-sonnet-4-20250514` |
34//! | `VERTEX_AI_PROJECT` | GCP project ID | required (or `GOOGLE_CLOUD_PROJECT`) |
35//! | `VERTEX_AI_LOCATION` | GCP region | `europe-west1` (or `GOOGLE_CLOUD_LOCATION`) |
36//! | `VERTEX_AI_TOKEN` | Static GCP auth token | auto-detect |
37//! | `VERTEX_AI_MODEL` | Vertex model name | `mistral-small-2503` |
38//! | `NOETHER_LLM_PROVIDER` | Force: `mistral` \| `openai` \| `anthropic` \| `vertex` \| `mock` | auto |
39//! | `NOETHER_EMBEDDING_PROVIDER` | Force: `mistral` \| `openai` \| `vertex` \| `mock` | auto |
40
41use crate::index::embedding::{EmbeddingProvider, MockEmbeddingProvider};
42use crate::llm::anthropic::AnthropicProvider;
43use crate::llm::mistral::{MistralNativeEmbeddingProvider, MistralNativeProvider};
44use crate::llm::openai::{OpenAiEmbeddingProvider, OpenAiProvider};
45use crate::llm::vertex::{
46    MistralLlmProvider, VertexAiConfig, VertexAiEmbeddingProvider, VertexAiLlmProvider,
47};
48use crate::llm::{LlmProvider, MockLlmProvider};
49
50/// Returns true if the model name is a Mistral/Codestral model.
51fn is_mistral_model(model: &str) -> bool {
52    let lower = model.to_lowercase();
53    lower.contains("mistral") || lower.contains("codestral")
54}
55
56/// Build the best available LLM provider based on env config.
57///
58/// Priority: Mistral native → Vertex AI → Mock.
59pub fn build_llm_provider() -> (Box<dyn LlmProvider>, &'static str) {
60    let forced = std::env::var("NOETHER_LLM_PROVIDER").unwrap_or_default();
61
62    match forced.as_str() {
63        "mock" => return (Box::new(MockLlmProvider::new("{}")), "mock"),
64        "mistral" => match build_mistral_native_llm() {
65            Ok(p) => return (p, "mistral-native"),
66            Err(e) => {
67                eprintln!("Warning: Mistral native LLM unavailable: {e}. Falling back.");
68            }
69        },
70        "openai" => match build_openai_llm() {
71            Ok(p) => return (p, "openai"),
72            Err(e) => {
73                eprintln!("Warning: OpenAI LLM unavailable: {e}. Falling back.");
74            }
75        },
76        "anthropic" => match build_anthropic_llm() {
77            Ok(p) => return (p, "anthropic"),
78            Err(e) => {
79                eprintln!("Warning: Anthropic LLM unavailable: {e}. Falling back.");
80            }
81        },
82        "vertex" => match build_vertex_or_mistral_llm() {
83            Ok((p, name)) => return (p, name),
84            Err(e) => {
85                eprintln!("Warning: Vertex AI LLM unavailable: {e}. Falling back to mock.");
86                return (Box::new(MockLlmProvider::new("{}")), "mock");
87            }
88        },
89        _ => {} // auto-detect below
90    }
91
92    // Auto-detect: Mistral native → OpenAI → Anthropic → Vertex → mock.
93    if let Ok(p) = build_mistral_native_llm() {
94        return (p, "mistral-native");
95    }
96    if let Ok(p) = build_openai_llm() {
97        return (p, "openai");
98    }
99    if let Ok(p) = build_anthropic_llm() {
100        return (p, "anthropic");
101    }
102    if let Ok((p, name)) = build_vertex_or_mistral_llm() {
103        return (p, name);
104    }
105    eprintln!("Warning: No LLM provider configured. Using mock.");
106    eprintln!("  Set MISTRAL_API_KEY for the native Mistral API (recommended),");
107    eprintln!("  or set OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_APPLICATION_CREDENTIALS.");
108    (Box::new(MockLlmProvider::new("{}")), "mock")
109}
110
111/// Build the best available embedding provider based on env config.
112///
113/// Priority: Mistral native → Vertex AI → Mock.
114pub fn build_embedding_provider() -> (Box<dyn EmbeddingProvider>, &'static str) {
115    let forced = std::env::var("NOETHER_EMBEDDING_PROVIDER").unwrap_or_default();
116
117    match forced.as_str() {
118        "mock" => return (Box::new(MockEmbeddingProvider::new(128)), "mock"),
119        "mistral" => match MistralNativeEmbeddingProvider::from_env() {
120            Ok(p) => return (Box::new(p), "mistral-native"),
121            Err(e) => {
122                eprintln!("Warning: Mistral native embeddings unavailable: {e}. Falling back.");
123            }
124        },
125        "openai" => match build_openai_embedding() {
126            Ok(p) => return (p, "openai"),
127            Err(e) => {
128                eprintln!("Warning: OpenAI embeddings unavailable: {e}. Falling back.");
129            }
130        },
131        "vertex" => match build_vertex_embedding() {
132            Ok(p) => return (p, "vertex"),
133            Err(e) => {
134                eprintln!("Warning: Vertex AI embeddings unavailable: {e}. Falling back to mock.");
135                return (Box::new(MockEmbeddingProvider::new(128)), "mock");
136            }
137        },
138        _ => {} // auto-detect below
139    }
140
141    // Auto-detect: Mistral native → OpenAI → Vertex → mock.
142    if let Ok(p) = MistralNativeEmbeddingProvider::from_env() {
143        return (Box::new(p), "mistral-native");
144    }
145    if let Ok(p) = build_openai_embedding() {
146        return (p, "openai");
147    }
148    if let Ok(p) = build_vertex_embedding() {
149        return (p, "vertex");
150    }
151    (Box::new(MockEmbeddingProvider::new(128)), "mock")
152}
153
154// ── Internal builders ────────────────────────────────────────────────────────
155
156fn build_mistral_native_llm() -> Result<Box<dyn LlmProvider>, String> {
157    Ok(Box::new(MistralNativeProvider::from_env()?))
158}
159
160fn build_openai_llm() -> Result<Box<dyn LlmProvider>, String> {
161    Ok(Box::new(OpenAiProvider::from_env()?))
162}
163
164fn build_anthropic_llm() -> Result<Box<dyn LlmProvider>, String> {
165    Ok(Box::new(AnthropicProvider::from_env()?))
166}
167
168fn build_openai_embedding() -> Result<Box<dyn EmbeddingProvider>, String> {
169    Ok(Box::new(OpenAiEmbeddingProvider::from_env()?))
170}
171
172fn build_vertex_or_mistral_llm() -> Result<(Box<dyn LlmProvider>, &'static str), String> {
173    let model = std::env::var("VERTEX_AI_MODEL")
174        .unwrap_or_else(|_| crate::llm::LlmConfig::default().model.clone());
175    let config = VertexAiConfig::from_env()?;
176
177    if is_mistral_model(&model) {
178        Ok((Box::new(MistralLlmProvider::new(config)), "mistral-vertex"))
179    } else {
180        Ok((Box::new(VertexAiLlmProvider::new(config)), "vertex"))
181    }
182}
183
184fn build_vertex_embedding() -> Result<Box<dyn EmbeddingProvider>, String> {
185    let config = VertexAiConfig::from_env()?;
186    let model = std::env::var("VERTEX_AI_EMBEDDING_MODEL").ok();
187    let dimensions = std::env::var("VERTEX_AI_EMBEDDING_DIMENSIONS")
188        .ok()
189        .and_then(|s| s.parse().ok());
190    Ok(Box::new(VertexAiEmbeddingProvider::new(
191        config, model, dimensions,
192    )))
193}
noether_engine/providers.rs

noether_engine/
providers.rs
