Provider Architecture Guide

Architecture Overview

Providers come in two tiers behind one implementation trait:

Tier 1 — Catalog-only (zero code). An OpenAI-compatible endpoint that differs only in base URL, auth env var, and advertised capabilities/models. It has no Rust module: one static entry in src/core/providers/registry/catalog.rs fully describes it, and the factory builds an OpenAILikeProvider from that data at runtime. This is the primary path for new integrations.

Tier 2 — Code-based. A provider needing custom request/response transformation, custom auth signing, non-standard streaming, or rich model metadata lives in src/core/providers/<name>/, implements LLMProvider, and is wired into routing.

Routing does not use trait objects. Router deployments store the closed Provider enum (src/core/providers/mod.rs), which dispatches to concrete provider structs. LLMProvider (src/core/traits/provider/llmprovider/traitdefinition.rs) is the interface every variant implements — implementing the trait alone does not make a provider routeable; enum variant, dispatch arm, and factory wiring are crate-level changes. Trait objects appear only at the edges: the error mapper (Box<dyn ErrorMapper<ProviderError>>) and the streaming return type (Pin<Box<dyn Stream<Item = Result<ChatChunk, ProviderError>> + Send>>).

Enumerating Current Providers

Do not rely on memorized counts — they change often. Count from source:

# Tier 1: one def_chat()/def_local_chat() call per entry
# (each count includes the helper fn definition itself, so subtract 1)
grep -c 'def_chat(' src/core/providers/registry/catalog.rs
grep -c 'def_local_chat(' src/core/providers/registry/catalog.rs

# Tier 2: code-based provider modules (base/factory/macros/registry are infrastructure)
ls -d src/core/providers/*/ | grep -vE '/(base|factory|macros|registry)/'

Adding a Tier 1 Provider (Primary Path)

Two edits, nothing else:

// src/core/providers/registry/catalog.rs
def_chat(
    "myprovider",
    "My Provider",
    "https://api.myprovider.com/v1",
    "MYPROVIDER_API_KEY",
),

// src/core/providers/mod.rs — annotation comment alongside the other Tier 1 notes
// myprovider: Tier 1 -> registry/catalog.rs

At runtime createprovider (src/core/providers/factory/mod.rs) matches the selector against the catalog, builds an OpenAILikeConfig via ProviderDefinition::toopenailikeconfig, and constructs openailike::OpenAILikeProvider::newforcatalog(oaiconfig, def.capabilities) into the Provider::OpenAILike variant.

Keyless local servers use deflocalchat (AuthType::None, skipapikey = true):

def_local_chat("myrunner", "My Runner", "http://localhost:1234/v1"),

Each entry is a ProviderDefinition (src/core/providers/registry/definition.rs). Defaults from def_chat can be overridden with struct-update syntax:

• alternateauthenvvars — env vars checked after authenv_var (see together)
• model_prefix: Some("xai/") — selector/model prefix stripping (see xai)
• custom capabilities profile replacing the default

OPENAILIKECATALOG_CAPABILITIES (ChatCompletion, ChatCompletionStream, ToolCalling, FunctionCalling)

• name aliases resolve via canonicalcatalogname() in catalog.rs (e.g.

"zhipuai" -> "zhipu")

Registry API (src/core/providers/registry/, re-exported in registry/mod.rs): istier1provider(name), getdefinition(name), canonicalcatalogname(name), PROVIDERCATALOG.

Core Trait Definition: LLMProvider

There are no associated types. Every fallible method returns the unified ProviderError directly.

// src/core/traits/provider/llm_provider/trait_definition.rs
pub trait LLMProvider: Send + Sync + Debug + 'static {
    // ===== Required =====
    fn name(&self) -> &str;
    fn capabilities(&self) -> &'static [ProviderCapability];
    fn models(&self) -> &[ModelInfo];

    fn get_supported_openai_params(&self, model: &str) -> &'static [&'static str];
    async fn map_openai_params(
        &self,
        params: HashMap<String, Value>,
        model: &str,
    ) -> Result<HashMap<String, Value>, ProviderError>;
    async fn transform_request(
        &self,
        request: ChatRequest,
        context: RequestContext,
    ) -> Result<Value, ProviderError>;
    async fn transform_response(
        &self,
        raw_response: &[u8],
        model: &str,
        request_id: &str,
    ) -> Result<ChatResponse, ProviderError>;
    fn get_error_mapper(&self) -> Box<dyn ErrorMapper<ProviderError>>;
    async fn chat_completion(
        &self,
        request: ChatRequest,
        context: RequestContext,
    ) -> Result<ChatResponse, ProviderError>;
    async fn health_check(&self) -> HealthStatus;
    async fn calculate_cost(
        &self,
        model: &str,
        input_tokens: u32,
        output_tokens: u32,
    ) -> Result<f64, ProviderError>;

    // ===== Provided (default = not_supported / trivial) =====
    fn error_provider_name(&self) -> &'static str { "provider" }
    fn supports_capability(&self, capability: &ProviderCapability) -> bool {
        self.capabilities().contains(capability)
    }
    fn supports_model(&self, model: &str) -> bool {
        self.models().iter().any(|m| m.id == model)
    }
    // supports_tools / supports_streaming / supports_embeddings /
    // supports_image_generation delegate to supports_capability();
    // supports_vision() currently returns false.
    async fn chat_completion_stream(&self, request: ChatRequest, context: RequestContext)
        -> Result<Pin<Box<dyn Stream<Item = Result<ChatChunk, ProviderError>> + Send>>, ProviderError>;
    async fn embeddings(&self, request: EmbeddingRequest, context: RequestContext)
        -> Result<EmbeddingResponse, ProviderError>;
    async fn image_generation(&self, request: ImageGenerationRequest, context: RequestContext)
        -> Result<ImageGenerationResponse, ProviderError>;
    async fn audio_transcription(&self, request: TranscriptionRequest, context: RequestContext)
        -> Result<TranscriptionResponse, ProviderError>;
    async fn audio_translation(&self, request: TranslationRequest, context: RequestContext)
        -> Result<TranslationResponse, ProviderError>;
    async fn text_to_speech(&self, request: SpeechRequest, context: RequestContext)
        -> Result<SpeechResponse, ProviderError>;
    async fn get_average_latency(&self) -> Result<std::time::Duration, ProviderError>; // default 100ms
    async fn get_success_rate(&self) -> Result<f32, ProviderError>;                   // default 0.99
    async fn estimate_tokens(&self, text: &str) -> Result<u32, ProviderError>;        // len()/4
}

Optional dispatch methods (streaming, embeddings, images, audio) must still be gated by the matching ProviderCapability — route selection checks supports_capability() before calling them.

Tier 2: Code-Based Provider Pattern

Directory Structure

Real examples: src/core/providers/cloudflare/ (small) and src/core/providers/openai_like/.

src/core/providers/my_provider/
├── mod.rs           # Module exports, LLMProvider impl may live here too
├── config.rs        # Config struct (often via define_provider_config!)
├── provider.rs      # Provider struct + LLMProvider impl
├── model_info.rs    # Static model metadata
├── streaming.rs     # SSE parsing (optional)
└── error.rs         # Error helpers/mappers (optional; legacy name = ProviderError alias)

Configuration

BaseConfig (src/core/providers/base/config.rs) carries apikey, apibase, endpointaccess, timeout (secs), maxretries, headers, organization, apiversion, with env fallbacks ({PROVIDER}APIKEY, {PROVIDER}APIBASE, ...) via fromenv(provider) / for_provider(provider).

Implement the ProviderConfig trait (src/core/traits/provider/config.rs) — note the accessor names differ from BaseConfig's helpers:

pub trait ProviderConfig: Send + Sync + Clone + Debug + 'static {
    fn validate(&self) -> Result<(), String>;
    fn api_key(&self) -> Option<&str>;
    fn api_base(&self) -> Option<&str>;
    fn timeout(&self) -> std::time::Duration;
    fn max_retries(&self) -> u32;
    fn endpoint_access(&self) -> ProviderEndpointAccess { ProviderEndpointAccess::PublicOnly }
    fn use_ssrf_safe_client(&self) -> bool { false }
    // validate_standard(provider_name): shared key/timeout/retries checks
}

The defineproviderconfig! macro (exported from base/config.rs) generates the struct, builders (withapikey, withbaseurl, withtimeout), fromenv(), getapikey()/getapibase(), and the ProviderConfig impl in one call.

Provider Implementation

Modeled on src/core/providers/cloudflare/provider.rs:

use std::sync::Arc;
use crate::core::providers::base::{header, BaseConfig, GlobalPoolManager, HttpMethod};
use crate::core::providers::ProviderError;
use crate::core::traits::error_mapper::trait_def::ErrorMapper;
use crate::core::traits::error_mapper::DefaultErrorMapper;
use crate::core::traits::provider::llm_provider::trait_definition::LLMProvider;
use crate::core::traits::provider::ProviderConfig;

pub struct MyProvider {
    config: MyProviderConfig,
    pool_manager: Arc<GlobalPoolManager>,
    models: Vec<ModelInfo>,
}

impl MyProvider {
    pub fn new(config: MyProviderConfig) -> Result<Self, ProviderError> {
        config.validate()
            .map_err(|e| ProviderError::configuration(PROVIDER_NAME, e))?;
        let http_config = BaseConfig {
            api_key: config.api_key().map(str::to_owned),
            api_base: config.api_base().map(str::to_owned),
            endpoint_access: config.endpoint_access(),
            timeout: config.timeout().as_secs(),
            max_retries: config.max_retries(),
            ..BaseConfig::default()
        };
        Ok(Self {
            config,
            pool_manager: Arc::new(GlobalPoolManager::new_for_provider(
                PROVIDER_NAME,
                http_config,
            )?),
            models: load_models(),
        })
    }
}

impl LLMProvider for MyProvider {
    fn name(&self) -> &'static str { PROVIDER_NAME }
    fn error_provider_name(&self) -> &'static str { PROVIDER_NAME }

    async fn chat_completion(
        &self,
        request: ChatRequest,
        context: RequestContext,
    ) -> Result<ChatResponse, ProviderError> {
        let api_key = self.config.api_key()
            .ok_or_else(|| ProviderError::authentication(PROVIDER_NAME, "API key required"))?;
        let mut headers = vec![header("Authorization", format!("Bearer {api_key}"))];
        headers.push(header("Content-Type", "application/json".to_string()));

        // url/model/request_id derived from config + request (elided)
        let body = self.transform_request(request, context).await?;
        let response = self.pool_manager
            .execute_request(&url, HttpMethod::POST, headers, Some(body))
            .await?;

        let status = response.status();
        if !status.is_success() {
            let body_text = response.text().await
                .map_err(|e| ProviderError::network(PROVIDER_NAME, e.to_string()))?;
            return Err(self.get_error_mapper().map_http_error(status.as_u16(), &body_text));
        }
        let raw = response.bytes().await
            .map_err(|e| ProviderError::network(PROVIDER_NAME, e.to_string()))?;
        self.transform_response(&raw, &model, &request_id).await
    }

    fn get_error_mapper(&self) -> Box<dyn ErrorMapper<ProviderError>> {
        Box::new(DefaultErrorMapper)
    }
    // ... remaining required methods
}

Registration (routing) Requirements

Adding a Tier 2 provider means crate-level wiring, in this order of authority:

1. Add src/core/providers/<name>/, its module declaration, and a closed Provider

enum variant under the same feature gate in src/core/providers/mod.rs.

1. Add the variant to all four dispatch_provider! @expand arms (sync,

asyncerr, value, asyncdirect) and to Provider::name() and provider_type().

1. Add a ProviderType variant and allnoncustomprovidertypes() entry in

provider_type.rs.

1. Add canonical name, aliases, catalog_backed and the correct

ProviderDispatchKind to PROVIDERTYPEREGISTRY in registry/types.rs. For a feature-gated native implementation, use or add a cfg-sensitive dispatch-kind helper that reports the correct enabled and disabled modes; registry entries themselves do not have a feature field.

1. Add the config builder in factory/builder.rs and match branch in

factory/registry.rs, with matching feature gates. Keep cfg gates synchronized across module, Provider, dispatch and factory wiring.

1. Update provider-type/registry lifecycle, factory support and feature-on/off

tests so aliases, support state and construction stay in sync.

Connection Pooling

The recommended newforprovider path runs requests through BaseHttpClient and ProviderHttpClient. ProviderHttpClient keeps a process-wide client cache keyed by endpoint policy, request timeout, and mode (ordinary, streaming, or no-redirect). Those clients use HttpClientPoolConfig::default() (src/utils/net/http.rs): 100 maximum idle connections per host, 90-second idle timeout, 10-second connect timeout, and 60-second TCP keepalive.

PoolConfig in src/core/providers/base/connection_pool.rs (there is no pool.rs) instead configures the legacy global ConnectionPool used by the unbound new() / shared() path. These constants do not tune the policy-bound clients above:

pub struct PoolConfig;
impl PoolConfig {
    pub const TIMEOUT_SECS: u64 = 600;
    pub const POOL_SIZE: usize = 80;      // pool_max_idle_per_host
    pub const KEEPALIVE_SECS: u64 = 90;   // pool_idle_timeout
}

// Provider implementations use new_for_provider(provider, BaseConfig) so
// endpoint-access and timeout policy are installed. new()/shared() omit that policy.
pub async fn execute_request(
    &self,
    url: &str,
    method: HttpMethod,               // GET | POST | PUT | DELETE
    headers: Vec<HeaderPair>,         // (Cow<'static, str>, Cow<'static, str>)
    body: Option<serde_json::Value>,  // serialized as JSON; not generic
) -> Result<reqwest::Response, ProviderError>;

pub async fn execute_streaming_request(
    &self,
    url: &str,
    headers: Vec<HeaderPair>,
    body: serde_json::Value,
    legacy_provider: &'static str,
) -> Result<reqwest::Response, ProviderError>;

Build header pairs zero-copy where possible: header(key, value) (static key, owned value), headerstatic(key, value), headerowned(key, value). Streaming callers get a separate client without a total-body timeout via streamingunboundedclient(); header phase bounded by STREAMINGHEADERTIMEOUTSECS, error bodies by readstreamingerrorbody (10s / 64KiB caps).

Model Information

ModelInfo (src/core/types/model.rs) is a plain serializable struct with Default:

pub struct ModelInfo {
    pub id: String,
    pub name: String,
    pub provider: String,
    pub max_context_length: u32,
    pub max_output_length: Option<u32>,
    pub supports_streaming: bool,
    pub supports_tools: bool,
    pub supports_multimodal: bool,
    pub input_cost_per_1k_tokens: Option<f64>,
    pub output_cost_per_1k_tokens: Option<f64>,
    pub currency: String,
    pub capabilities: Vec<ProviderCapability>,
    pub created_at: Option<SystemTime>,
    pub updated_at: Option<SystemTime>,
    pub metadata: HashMap<String, serde_json::Value>,
}

Costs are per 1K tokens (catalog pricing constants in registry/catalog.rs are per million and divided by 1000 when converted). Return &self.models from LLMProvider::models(); supportsmodel() scans that slice.

Provider Capabilities

ProviderCapability (src/core/types/model.rs) — the full current variant list:

pub enum ProviderCapability {
    ChatCompletion,
    ChatCompletionStream,
    Embeddings,
    ImageGeneration,
    ImageEdit,
    ImageVariation,
    AudioTranscription,
    AudioTranslation,
    TextToSpeech,
    Moderation,
    Rerank,
    ToolCalling,
    FunctionCalling,
    CodeExecution,
    FileUpload,
    FineTuning,
    BatchProcessing,
    RealtimeApi,
    GeminiGenerateContent,
}

Return a static slice (a temporary array literal cannot back &'static [...]):

const MY_CAPABILITIES: &[ProviderCapability] = &[
    ProviderCapability::ChatCompletion,
    ProviderCapability::ChatCompletionStream,
    ProviderCapability::ToolCalling,
];

fn capabilities(&self) -> &'static [ProviderCapability] {
    MY_CAPABILITIES
}

Unified Error Handling

ProviderError (src/core/providers/unifiedprovidererror.rs, re-exported as crate::core::providers::ProviderError and unifiedprovider::ProviderError) is the single error type across the trait. Construct with factory methods (src/core/providers/unifiedprovider_methods.rs) rather than struct literals — some variants carry extra optional fields (e.g. RateLimit rpm/tpm limits):

ProviderError::authentication(provider, msg)
ProviderError::rate_limit(provider, retry_after: Option<u64>)
ProviderError::rate_limit_with_retry(provider, msg, retry_after)
ProviderError::model_not_found(provider, model)
ProviderError::invalid_request(provider, msg)
ProviderError::network(provider, msg)
ProviderError::timeout(provider, msg)
ProviderError::api_error(provider, status: u16, msg)
ProviderError::provider_unavailable(provider, msg)
ProviderError::not_supported(provider, feature)
ProviderError::configuration(provider, msg)
ProviderError::serialization(provider, msg)

ErrorMapper<E> (src/core/traits/errormapper/traitdef.rs) converts HTTP statuses and JSON error bodies into errors: required maphttperror(u16, &str); defaulted mapjsonerror, mapnetworkerror, mapparsingerror, maptimeouterror. Ready-made mappers:

• GenericErrorMapper — core::traits::error_mapper::types, aliased

DefaultErrorMapper at core::traits::errormapper (what most providers return from geterror_mapper())

• OpenAIErrorMapper, AnthropicErrorMapper — core::traits::error_mapper::implementations

Legacy per-provider error enums were removed; surviving names like AnthropicError or GeminiError are pub type ... = ProviderError aliases (see [reference/migration-from-legacy-errors.md](reference/migration-from-legacy-errors.md)).

References

• [reference/best-practices-and-checklist.md](reference/best-practices-and-checklist.md) — error factory conventions, coding practices, and new-provider checklist (Tier 1 and Tier 2)
• [reference/migration-from-legacy-errors.md](reference/migration-from-legacy-errors.md) — how legacy per-provider error types map onto unified ProviderError