Auto-Instrumentation

Auto-instrumentation patches LLM clients and agent frameworks at init time, capturing spans, token counts, latency, prompts, and responses without any changes to your application code.

How It Works

Pass the libraries you use in the instrumentations list when calling neatlogs.init():

neatlogs.init(
    api_key=os.environ["NEATLOGS_API_KEY"],
    endpoint=os.environ["NEATLOGS_ENDPOINT"],
    workflow_name="my-app",
    instrumentations=["openai", "langchain", "chromadb"],
)

The SDK patches each library's internals at init time. Any call made after init() is automatically traced — no decorators or context managers required for the instrumented libraries themselves.

The init() Placement Rule

neatlogs.init() must be called before any instrumented library is imported or used. Instrumentation patches library internals at the point init() is called.

# CORRECT
import neatlogs
neatlogs.init(
    api_key=...,
    endpoint=...,
    instrumentations=["langchain", "openai"],
)
from langchain_openai import ChatOpenAI   # patched correctly

# WRONG
from langchain_openai import ChatOpenAI
import neatlogs
neatlogs.init(...)    # too late — LangChain will not be instrumented

For multi-file projects, place init() in your entrypoint (main.py) before importing your workflow modules:

# main.py
import neatlogs

neatlogs.init(
    api_key=os.environ["NEATLOGS_API_KEY"],
    endpoint=os.environ["NEATLOGS_ENDPOINT"],
    instrumentations=["langchain", "openai", "crewai"],
)

from workflows.customer_support import run_customer_support
from workflows.research import run_research

run_customer_support()
run_research()

neatlogs.flush()
neatlogs.shutdown()

The TracerProvider registered by init() is global — any module imported after init() will have its LLM and framework calls captured automatically.

Span Deduplication

Neatlogs uses two instrumentation layers under the hood (OpenLLMetry and OpenInference). Some libraries are instrumented by both, which can produce duplicate spans. The SDK's NeatlogsSpanProcessor detects and deduplicates overlapping spans automatically — you will not see duplicates in the dashboard.

What Gets Captured

For each instrumented LLM call:

• Model name
• Prompt messages and completion response
• Token counts (prompt, completion, total, cache)
• Latency

For each instrumented framework operation (chains, agents, tools):

• Span kind (inferred from the library)
• Input and output
• Latency

See Supported Libraries for the full list of valid instrumentations keys.