Observability with Langfuse

Langfuse  is an open-source observability and analytics platform for LLM-powered applications. It collects traces, generations, and custom metrics so that you can debug, evaluate, and monitor your AI features in production.

Integration inside Cognee

Cognee ships with Langfuse support out of the box. The integration is intentionally lightweight and consists of just a few lines of code. Anywhere in the codebase where we want observability we add:

from cognee.modules.observability.get_observe import get_observe
observe = get_observe()
 
@observe(as_type="generation")  # optional label
async def acreate_structured_output(...):
    ...

The decorator is a thin wrapper around Langfuse and automatically creates a span every time the function is executed.

The data is sent to the Langfuse backend specified via environment variables and can be inspected in the Langfuse UI.

Quick start

1. Install cognee (it comes with Langfuse)

   Langfuse is already declared in pyproject.toml, so you can skip this step if you already installed cognee.

2. Create a project at Langfuse Cloud  and copy the public and secret keys.

3. Export the following environment variables (for example in your .env file):

LANGFUSE_PUBLIC_KEY=<your public key>
LANGFUSE_SECRET_KEY=<your secret key>
LANGFUSE_HOST=https://cloud.langfuse.com

BaseConfig reads these values on startup, so nothing else is required. Start Cognee as usual and open the Langfuse UI – traces will appear in real time.

Run your regular Cognee workflows:

import cognee
import asyncio
from cognee.modules.observability.get_observe import get_observe
 
observe = get_observe()
 
@observe(name="simple_example_run", as_type="example")
async def main():
    await cognee.add("Natural language processing (NLP) is ...")
    await cognee.cognify()
    results = await cognee.search("Tell me about NLP")
    for r in results:
        print(r)
 
asyncio.run(main())

Adding your own spans

You can instrument any function in your codebase:

from cognee.modules.observability.get_observe import get_observe
observe = get_observe()
 
@observe(as_type="my_tool", metadata={"foo": "bar"})
def my_helper(arg1, arg2):
    ...

All keyword arguments accepted by langfuse.decorators.observe (name, as_type, metadata, …) are forwarded unchanged – see the Langfuse Python SDK docs  for the full reference.

Join Langfuse  and cognee  communities on Discord for any questions.