Skip to main content
This page documents the public surface of trulayer. For narrative usage, see the tutorial. For configuration details, see configuration.

​
Module functions

​
trulayer.init()

Initialise the global client. Call once at app startup. 
trulayer.init(
    api_key: str,
    project: str,
    environment: str = "production",
    endpoint: str = "https://api.trulayer.ai",
    batch_size: int = 50,
    flush_interval: float = 2.0,
    timeout: float = 5.0,
    sample_rate: float = 1.0,
    scrub_fn: Callable[[Any], Any] | None = None,
    metadata_validator: Callable[[dict], bool] | None = None,
    debug: bool = False,
) -> TruLayerClient
Returns the created TruLayerClient. Stored as the process-wide global — retrievable via get_client().

​
trulayer.get_client()

trulayer.get_client() -> TruLayerClient
Return the global client. Raises RuntimeError if init() has not been called.

​
trulayer.trace()

Context manager that starts a new trace. 
trulayer.trace(
    name: str,
    session_id: str | None = None,
    metadata: dict | None = None,
) -> TraceContext
Usage: 
with trulayer.trace("answer_question") as trace:
    trace.set_input({"question": question})
    ...
    trace.set_output({"answer": answer})

​
trulayer.atrace()

Async variant of trace(). Same signature; used with async with. 
async with trulayer.atrace("answer_question") as trace:
    trace.set_input({"question": question})

​
trulayer.current_trace()

trulayer.current_trace() -> TraceContext | None
Return the trace in the current async-local context, or None if no trace is active.

​
trulayer.shutdown()

Flush buffered spans and stop the background worker. Call on process exit. 
trulayer.shutdown()

​
Instrumentation helpers

​
trulayer.instrument_openai()

trulayer.instrument_openai(client: openai.OpenAI | openai.AsyncOpenAI) -> None
Patch client.chat.completions.create and client.embeddings.create to emit spans automatically. Reversible via uninstrument_openai(client).

​
trulayer.instrument_anthropic()

trulayer.instrument_anthropic(client: anthropic.Anthropic | anthropic.AsyncAnthropic) -> None
Patch client.messages.create to emit spans. Reversible via uninstrument_anthropic(client).

​
trulayer.instrument_langchain()

trulayer.instrument_langchain() -> BaseCallbackHandler
Return a LangChain callback handler. Pass it to any ChatModel, Chain, or Retriever via the callbacks=[...] argument.

​
Classes

​
TruLayerClient

Explicit client for multi-tenant apps or tests. Prefer init() + get_client() for single-client apps. 
from trulayer import TruLayerClient

client = TruLayerClient(api_key="...", project="...")
with client.trace("work") as trace:
    ...
Key methods:
MethodSignature
trace(name, session_id=None, metadata=None)→ TraceContext
atrace(name, ...)→ async TraceContext
submit_feedback(trace_id, label=None, score=None, comment=None, metadata=None)→ None
run_eval(trace_id, evaluator_type, metric_name)→ EvalResult
get_metrics(project_id, from_time, to_time, ...)→ Metrics
flush()→ None (blocks until shipped)
shutdown()→ None

​
TraceContext

Returned by trace() and atrace(). Key methods:
MethodPurpose
set_input(value)Set the trace’s input payload
set_output(value)Set the trace’s output payload
set_metadata(**kwargs)Attach key-value metadata
set_error(exc)Mark the trace as errored with an exception
span(name, span_type="custom", ...)→ SpanContext
.idTrace UUID (read-only)
.session_idSession identifier (read-only)

​
SpanContext

Returned by TraceContext.span(). Key methods:
MethodPurpose
set_input(value)Span input
set_output(value)Span output
set_metadata(**kwargs)Key-value metadata
set_model(model)Model name (for llm spans)
set_tokens(prompt_tokens, completion_tokens)Token counts
set_error(exc)Mark errored
.idSpan UUID (read-only)

​
Models (Pydantic)

​
TraceData

class TraceData(BaseModel):
    id: UUID
    project_id: str
    session_id: str | None
    name: str
    input: Any
    output: Any
    error: ErrorInfo | None
    tags: list[str]
    metadata: dict[str, Any]
    spans: list[SpanData]
    started_at: datetime
    ended_at: datetime | None

​
SpanData

class SpanData(BaseModel):
    id: UUID
    trace_id: UUID
    name: str
    span_type: Literal["llm", "retrieval", "tool", "custom"]
    input: Any
    output: Any
    error: ErrorInfo | None
    latency_ms: int
    model: str | None
    prompt_tokens: int | None
    completion_tokens: int | None
    metadata: dict[str, Any]
    started_at: datetime
    ended_at: datetime | None

​
FeedbackData

class FeedbackData(BaseModel):
    trace_id: UUID
    label: Literal["good", "bad", "neutral"] | None
    score: float | None
    comment: str | None
    metadata: dict[str, Any]

​
EventData

class EventData(BaseModel):
    id: UUID
    trace_id: UUID
    span_id: UUID | None
    level: Literal["debug", "info", "warn", "error"]
    message: str
    metadata: dict[str, Any]
    timestamp: datetime

​
Exceptions

ExceptionRaised when
TruLayerErrorBase class for all SDK errors
AuthenticationErrorInvalid or revoked API key
RateLimitErrorPlan limit hit; trace dropped
ValidationErrorTrace/span payload failed validation
The SDK never raises these to your call site — they’re logged and the trace is dropped. Subscribe via trulayer.on_error(fn) if you need to handle them yourself.