Opik

Opik SDK Reference

Opik is an open-source LLM observability platform. This skill is a reference for the SDK. To instrument a codebase step by step (detect frameworks, add config, emit and verify a trace), use the task-shaped opik-instrument skill.

Core concepts

A trace is one execution path (one request → one response). Spans are the operations inside it and form a hierarchy.

Span types — the ONLY valid values

Do NOT use retrieval or any other value.

Python — tracing

import opik

@opik.track(name="agent", type="general")
def agent(query: str) -> str:
    return generate(retrieve(query))

@opik.track(type="tool")
def retrieve(query): ...

@opik.track(type="llm")
def generate(ctx): ...

opik.flush_tracker()   # required in scripts

TypeScript — tracing

import { Opik } from "opik";
const client = new Opik({ projectName: "my-project" });

const trace = client.trace({ name: "agent", input: { query } });
const span = trace.span({ name: "llm-call", type: "llm" });
span.end({ output });
trace.end({ output });
await client.flush();

Framework integrations

Prefer an integration over manual @opik.track — integrations capture tokens, model, and cost automatically. Patterns (full list in references/integrations.md):

• wrap-the-client — track_openai(OpenAI()), track_anthropic(...)
• global-enable — track_crewai(crew=crew)
• callback — dspy.configure(callbacks=[OpikCallback()])
• tracer — OpikTracer() for LangChain / LangGraph / LlamaIndex
• agent-specific — track_adk_agent_recursive(agent, OpikTracer())

LiteLLM inside @opik.track (common trap)

If code uses litellm and you add @opik.track, pass current_span_data via metadata on every completion call — otherwise OpikLogger emits orphaned top-level traces instead of nesting under your span.

from opik.opik_context import get_current_span_data

@opik.track
def call_llm(messages):
    return litellm.completion(
        model="gpt-4o", messages=messages,
        metadata={"opik": {"current_span_data": get_current_span_data()}},
    )

Threads (conversations)

Group turns with thread_id — one turn = one trace, shared thread_id = one thread. Use for chat / multi-turn; skip for single-shot.

@opik.track(entrypoint=True)
def handle(session_id: str, message: str) -> str:
    opik.update_current_trace(thread_id=session_id)
    return reply(message)

Prompt library

Version prompts with client.get_prompt / create_prompt (chat variants: get_chat_prompt / create_chat_prompt). Store model + temperature in the prompt metadata so they version with the text. Call get_prompt inside a @opik.track function so the version links to the trace.

@opik.track(entrypoint=True)
def run(question: str) -> str:
    p = client.get_prompt(name="system") or client.create_prompt(
        name="system",
        prompt="You help with {{product}}.",
        metadata={"model": "gpt-4o", "temperature": 0.7},
    )
    return llm(p.format(product="Opik"), model=p.metadata["model"])

Searching traces

One filter grammar, OQL, serves both the hosted MCP's list tool and the SDK's search_traces / search_spans / search_threads:

<field>[.<key>] <op> <value> [AND ...]
ops: = != > >= < <= contains not_contains starts_with ends_with is_empty is_not_empty in not_in

Strings in double quotes, numbers bare, duration in milliseconds, dates as ISO-8601 instants with a timezone ("2026-09-08T10:00:00Z"). Scores and dictionaries take a key: feedback_scores.accuracy < 0.5, metadata.environment = "prod". AND is the only connector.

error_info is_not_empty AND duration > 5000
type = "llm" AND usage.total_tokens > 10000            # spans
feedback_scores.hallucination > 0.5 AND start_time >= "2026-09-08T00:00:00Z"

With the MCP connected, prefer list — it also sorts (sort="duration desc"), windows (since="1h", "7d"), and searches free text (search="order-42"):

list(entity_type="trace", project_name="<project>", since="1h",
     filters="error_info is_not_empty", sort="duration desc")

Trace, span and thread lists add source = "sdk" unless you name source, so evaluator, playground and experiment traces stay out of the way. A rejected filter comes back with what fixes it; schema("list.trace") (or list.span, list.thread, list.experiment) is the full field and operator reference.

Without the MCP, the same string goes to the SDK:

client.search_traces(project_name="<project>", filter_string="error_info is_not_empty")

Anti-patterns

References

To build and run an evaluation, use the opik-evaluate skill. For repo instrumentation and config, use the opik-instrument skill.