Agent observability: LangSmith, Helicone, Phoenix Arize

# Setup: variáveis de ambiente + uso automático
import os
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_API_KEY"] = "lsv2_..."
os.environ["LANGCHAIN_PROJECT"] = "ffv-agente-prod"

from langchain_anthropic import ChatAnthropic
from langgraph.graph import StateGraph

# Toda call LLM e nó LangGraph é tracked automaticamente
llm = ChatAnthropic(model="claude-sonnet-4-7")
response = llm.invoke("Olá")
# → aparece em smith.langchain.com com latência, tokens, custo, prompt, response

# Decorators para custom code
from langsmith import traceable

@traceable(name="custom-retrieval", run_type="retriever")
def my_retrieval(query: str) -> list[str]:
    # Sua lógica de retrieval — vira span filho do trace pai
    return search_db(query)

# Prompt versioning via Hub
from langchain import hub
prompt = hub.pull("my-team/research-agent:v3")  # versão pinada
response = llm.invoke(prompt.format(question="..."))

# Setup: muda base_url para Helicone — não exige callbacks/framework
import openai
import anthropic

# OpenAI via Helicone
openai_client = openai.OpenAI(
    base_url="https://oai.helicone.ai/v1",
    default_headers={
        "Helicone-Auth": f"Bearer {os.environ['HELICONE_API_KEY']}",
        "Helicone-User-Id": "user-42",
        "Helicone-Property-Feature": "research",      # custom properties
    },
)

# Anthropic via Helicone
anthropic_client = anthropic.Anthropic(
    base_url="https://anthropic.helicone.ai",
    default_headers={
        "Helicone-Auth": f"Bearer {os.environ['HELICONE_API_KEY']}",
        "Helicone-Cache-Enabled": "true",             # cache built-in
        "Helicone-Rate-Limit-Policy": "1000;w=60",    # 1000 req/min
    },
)

# Toda call passa pelo proxy — captura automática + features extras:
# - Cache de responses repetidos
# - Rate limiting per user
# - Cost tracking per property
# - Prompt management via Helicone Prompts

# Phoenix self-host (Docker) + auto-instrumentation via openinference
import phoenix as px
from phoenix.otel import register
from openinference.instrumentation.openai import OpenAIInstrumentor
from openinference.instrumentation.langchain import LangChainInstrumentor

# Inicia UI local ou aponta para servidor remoto
px.launch_app()  # http://localhost:6006

# Configura tracer OTel — exports para Phoenix ou qualquer OTel backend
tracer_provider = register(
    project_name="ffv-agente",
    endpoint="http://localhost:6006/v1/traces",
)

# Auto-instrument SDKs
OpenAIInstrumentor().instrument(tracer_provider=tracer_provider)
LangChainInstrumentor().instrument(tracer_provider=tracer_provider)

# Specialty: embedding visualizations
# Phoenix mostra UMAP/t-SNE de embeddings, detecta drift entre prod/eval datasets
# Ideal para RAG observability — visualiza queries que retornam docs irrelevantes

from langfuse import Langfuse
from langfuse.decorators import observe

langfuse = Langfuse(
    public_key=os.environ["LANGFUSE_PUBLIC_KEY"],
    secret_key=os.environ["LANGFUSE_SECRET_KEY"],
    host="https://cloud.langfuse.com",   # ou self-host
)

@observe()
def research_agent(query: str) -> str:
    # Trace é criado automaticamente — child spans capturam calls LLM internas
    prompt = langfuse.get_prompt("research-system", version=3)
    messages = [{"role": "system", "content": prompt.compile()}]
    messages.append({"role": "user", "content": query})

    response = anthropic_client.messages.create(
        model="claude-sonnet-4-7",
        messages=messages,
        max_tokens=2000,
    )

    # Captura score (eval feedback ou LLM-as-judge)
    langfuse.score(
        trace_id=langfuse.get_current_trace_id(),
        name="relevance",
        value=0.85,
        comment="LLM-as-judge: relevância da resposta",
    )

    return response.content[0].text