llm-obs 0.1.0

Lightweight SDK for LLM inference logging and observability

llm-obs

Lightweight Python SDK for LLM inference logging and observability.

Auto-instruments OpenAI, Anthropic, Google Gemini, AWS Bedrock, Ollama, and any OpenAI-compatible endpoint — zero changes to your LLM call code.

---

Install

# Core SDK
pip install llm-obs

# With provider extras
pip install "llm-obs[openai]"
pip install "llm-obs[anthropic]"
pip install "llm-obs[gemini]"
pip install "llm-obs[bedrock]"
pip install "llm-obs[all]"

---

Quickstart — one line

from llm_obs import ObservabilityClient

obs = ObservabilityClient(
    endpoint="http://localhost:4000",   # your ingestion API
    api_key="dev-key",
)
obs.auto_instrument()   # patches all installed LLM libraries automatically

From this point, every LLM call in your app is logged automatically. No other changes needed.

---

Stream chat

from llm_obs import stream_chat, set_obs_context

# Set conversation context (picked up automatically by the SDK)
set_obs_context(conversation_id="conv-123")

# Unified streaming across all providers
async for chunk in stream_chat(provider="openai", model="gpt-4o-mini", messages=[
    {"role": "user", "content": "Explain Redis in one sentence."}
]):
    print(chunk, end="", flush=True)

---

Provider detection from URL

from llm_obs import detect_provider, available_providers
import os

os.environ["LLM_ENDPOINTS"] = "http://localhost:11434"  # Ollama, vLLM, or any URL

# SDK probes the URL and detects what's running
providers = available_providers()
# → {"ollama": ["gemma3:4b", "llama3.2", ...]}

Supported URL detection:

• Ollama — detected via GET /api/tags
• vLLM / LiteLLM / LocalAI — detected via GET /v1/models
• AWS Bedrock — detected from URL pattern (amazonaws.com)
• OpenAI / Anthropic / Google — detected from known API URL patterns
• Private VPC — probed automatically

---

What gets logged per call

Field	Description
provider / model	Who served the request
latency_ms	Total wall-clock time
ttft_ms	Time-to-first-token (streaming)
prompt_tokens / completion_tokens	Token usage
cost_usd	Computed from built-in price table
status	success, error, cancelled
request / response	PII-redacted payloads
conversation_id	Linked via set_obs_context()

---

PII redaction

PII is redacted in-process before data leaves via HTTP — email, phone, SSN, credit cards (Luhn), API keys, IPv4, URL secrets.

obs = ObservabilityClient(..., redact_pii=True)   # default: True

---

Manual span

span = obs.start_span(
    provider="openai",
    model="gpt-4o-mini",
    request={"messages": [{"role": "user", "content": "Hello"}]},
    conversation_id="conv-123",
)
span.set_ttft(ms=210)
span.set_usage(prompt_tokens=42, completion_tokens=11)
span.end(status="success", streamed=True)

---

License

MIT