Crate synth_ai

Expand description

§Synth AI SDK

Ergonomic Rust SDK for Synth AI - serverless post-training APIs.

§Quick Start

use synth_ai::Synth;

#[tokio::main]
async fn main() -> Result<(), synth_ai::Error> {
    // Create client from SYNTH_API_KEY env var
    let synth = Synth::from_env()?;

    // Submit a prompt optimization job
    let result = synth
        .optimize()
        .task_app("https://my-task-app.com")
        .model("gpt-4o")
        .run()
        .await?;

    println!("Best prompt: {:?}", result.best_prompt);
    Ok(())
}

Re-exports§

pub use environment_pools::EnvironmentPoolsClient;
pub use synth_ai_core as core;
pub use synth_ai_core_types as types;

Modules§

environment_pools

Structs§

Artifact: An artifact produced during a rollout or evaluation.
BufferedHandler: A handler that buffers messages in memory.
CalibrationExample: Calibration example for verifier evaluation.
CallbackHandler: A handler that calls a callback function.
CandidateInfo: Information about a single candidate.
ContextOverride: A context override containing file artifacts, env vars, and scripts.
ContextOverrideStatus: Result of applying a context override.
CoreClient: Synth API client.
Criterion: A single evaluation criterion.
CriterionScoreData: Score data for a single criterion.
EvalBuilder: Builder for evaluation jobs.
EvalJobRequest: Request to submit an evaluation job.
EventObjectiveAssignment: Assignment of objectives to an event.
EventRewardRecord: Event-level reward annotation.
GEPAProgress: Overall GEPA progress state.
GepaJobRequest: Request to submit a GEPA optimization job.
GoldExample: Gold-standard example for contrastive evaluation.
GraphCompletionRequest: Request for a graph completion.
GraphCompletionResponse: Response from a graph completion.
GraphEvolveClient: Graph Evolve API client.
GraphEvolveJob: High-level Graph Evolve job orchestration.
InstanceObjectiveAssignment: Assignment of objectives to a specific instance (dataset example).
JobStreamer: Job streamer that polls endpoints and dispatches to handlers.
JsonHandler: A handler that outputs JSON lines.
Judgement: A complete judgement including rubric assignment and annotations.
LLMCallRecord: Normalized LLM call record.
LLMChunk: Optional streaming chunk representation.
LLMContentPart: LLM message content part.
LLMMessage: LLM message.
LLMRequestParams: Provider request parameters.
LLMUsage: Token usage statistics.
LibsqlTraceStorage: libsql-based trace storage.
LocalApiDeployResponse: Response for a LocalAPI deployment request.
LocalApiDeploySpec: Deployment specification for managed LocalAPI builds.
LocalApiDeployStatus: Status response for LocalAPI deployments.
LocalApiDeploymentInfo: Deployment list entry for LocalAPI deployments.
LocalApiLimits: Resource limits for managed LocalAPI deployments.
MarkovBlanketMessage: Inter-system message (Markov blanket).
MessageContent: Content for messages, supporting text or JSON.
MiproJobRequest: Request to submit a MIPRO optimization job.
ObjectiveSpec: Specification for an optimization objective.
OptimizeBuilder: Builder for prompt optimization jobs.
OptimizeResult: Result of a prompt optimization job.
OutcomeObjectiveAssignment: Assignment of objectives to an outcome (session-level).
OutcomeRewardRecord: Episode-level reward summary.
ProgressTracker: Progress tracker that aggregates events into state.
PromptLearningJob: High-level prompt learning job orchestration.
PromptLearningResult: Result of a policy optimization job (GEPA/MIPRO).
PromptResults: Extracted prompt results.
RankedPrompt: Ranked prompt from results.
RewardAggregates: Aggregated statistics for rewards.
RewardObservation: A single reward observation.
RlmOptions: Options for RLM (Retrieval-augmented LM) inference.
Rubric: A rubric containing multiple evaluation criteria.
RubricAssignment: Assignment of scores to a rubric’s criteria.
SessionTimeStep: A timestep within a session.
SessionTrace: A complete session trace.
SessionTracer: Session tracer for recording traces.
StreamConfig: Configuration for stream filtering and behavior.
StreamEndpoints: Endpoint configuration for streaming from a job.
StreamMessage: A message from a job stream.
Synth: Main Synth AI client.
TaskAppClient: Client for communicating with task apps.
TimeRecord: Time record for events and messages.
ToolCallResult: Tool call result.
ToolCallSpec: Tool call specification.
TraceUploadClient: Client for trace upload endpoints.
TunnelHandle
UploadUrlResponse: Response from creating an upload URL.
VerifierOptions: Options for verifier inference.
VerifierResponse: Response from verifier inference.

Enums§

ApplicationErrorType: Type of error during override application.
ApplicationStatus: Status of a context override application.
CoreError: Unified error enum for all Synth core errors.
Error: SDK error type.
EvalJobStatus: Status for evaluation jobs.
JobType: Type of optimization/training job.
PolicyJobStatus: Status for policy optimization jobs (GEPA, MIPRO).
StreamType: Type of stream data.
SynthModelName: Synth-hosted model names.
TracingEvent: Unified event type using tagged enum.
TunnelBackend
TunnelError

Constants§

API_KEY_ENV: Environment variable for API key.
DEFAULT_BASE_URL: Default Synth API base URL.
VERSION: SDK version.

Traits§

StreamHandler: Trait for handling stream messages.

Functions§

eval: Create a client from environment and run a quick evaluation.
open_tunnel
optimize: Create a client from environment and run a quick optimization.

Type Aliases§

Result: Result type alias.

Crate synth_ai

Crate synth_ai Copy item path

§Synth AI SDK

§Quick Start

Re-exports§

Modules§

Structs§

Enums§

Constants§

Traits§

Functions§

Type Aliases§

Crate synth_ai