An observability system makes production LLM behavior visible when healthy logs and strong development results fail to explain confusing or inconsistent user responses. It helps teams investigate whether problems stem from prompts, model drift, context length, or retrieval. section 10.6
For LLM applications, useful visibility can include:
- End-to-end traces of user input, retrieval, and model calls. section 10.6
- Token counts, costs, and latency for pipeline components. section 10.6
- Prompt and model versions, output-quality scores, errors, retries, and fallbacks. section 10.6
- Quality and reliability signals such as semantic quality, hallucination rates, user satisfaction, and drift. section 1.4.3
Résumé generation
Huntr adopted Langfuse to track LLM generations and support quality assurance and debugging for its AI Résumé Builder. The example instrumentation wraps résumé generation in a trace and records a quality score. section 10.6 section 10.6
from langfuse import observe, get_client
from langfuse.openai import openai
langfuse = get_client()
@observe(name="resume-generator")
def build_resume(user_input):
return openai.chat.completions.create(
model="gpt-5.4",
messages=[
{"role": "system", "content": "You are an expert resume
writer."},
{"role": "user", "content": user_input}
]
).choices[0].message.content
with langfuse.start_as_current_span(name="resume-session",
metadata={"user_id": "user_1234"}) as span:
response = build_resume("I worked at Stripe as a backend engineer for 3
years...")
span.score_trace(name="resume_quality_score", value=4)
The Python example instruments a résumé-generation call with a trace, session metadata, and a quality score. section 10.6
