emberlm

Official Python SDK for EmberLM. Run versioned prompts from your production code, evaluate LLM responses, and ship AI with confidence.

Install

pip install emberlm

Requires Python 3.8+. Zero third-party dependencies.

Quick start

Get an API key from Settings → API Keys in the dashboard. Keys start with pk_live_.

from emberlm import Client

client = Client(api_key="pk_live_...")

result = client.run(
    "summarize_docs",
    variables={"document": doc_text},
)

print(result["text"])
print("passed_evals:", result["passed_evals"])
print("confidence:", result["confidence"])
print("cost_usd:", result["cost_usd"])
print("latency_ms:", result["latency_ms"])

API

Client(api_key, base_url=..., timeout=60)

client = Client(
    api_key="pk_live_...",
    base_url="https://emberlm.dev",  # optional
    timeout=60,                        # optional, seconds
)

client.list_prompts()

List every prompt in the API key's workspace.

data = client.list_prompts()
for p in data["prompts"]:
    print(p["name"], p["current_version"])

client.get_prompt(name)

Fetch a single prompt by name. Returns the prod-tagged version if one exists, otherwise the latest.

prompt = client.get_prompt("summarize_docs")
print(prompt["system_prompt"], prompt["user_prompt"])

client.run(prompt_name, variables=None, model=None)

Run a saved prompt. Variables are substituted into {{placeholders}}. All workspace eval rules are applied to the output. The run is persisted and counts toward the workspace's analytics.

result = client.run(
    "classify_ticket",
    variables={"body": ticket_body},
    model="claude-haiku-4-5",  # optional override
)

# result keys:
#   run_id, prompt, version, model,
#   text,
#   input_tokens, output_tokens, total_tokens,
#   cost_usd, latency_ms,
#   passed_evals, confidence, evals,
#   error

client.eval(response, rule_ids=None, prompt=None, variables=None)

Evaluate an arbitrary response against the workspace's active eval rules. Useful when the response was generated by a different SDK or model.

result = client.eval(
    response=llm_response,
    prompt=user_prompt,            # optional
    variables={"name": "Jane"},    # optional
    rule_ids=["rule-id-1"],        # optional, defaults to all active
)
print(result["passed"], result["confidence"])

An alias client.evaluate(...) is also available for users who prefer to avoid the eval name.

Errors

Any non-2xx response raises EmberLMError:

from emberlm import Client, EmberLMError

try:
    client.run("missing_prompt")
except EmberLMError as e:
    print(e.status, str(e))

Rate limits

100 requests / minute per API key.

License

MIT