LangSmith va Weights & Biases

1import os
2os.environ["LANGCHAIN_TRACING_V2"] = "true"
3os.environ["LANGCHAIN_API_KEY"] = "lsv2_pt_xxx"
4os.environ["LANGCHAIN_PROJECT"] = "my-ai-project"
5
6# Moi LangChain call se tu dong duoc traced
7from langchain_openai import ChatOpenAI
8from langchain_core.prompts import ChatPromptTemplate
9
10llm = ChatOpenAI(model="gpt-4o-mini")
11chain = ChatPromptTemplate.from_messages([
12    ("system", "Ban la AI assistant."),
13    ("human", "{question}")
14]) | llm
15
16# This call is automatically traced in LangSmith
17result = chain.invoke({"question": "Hello!"})

1from langsmith import traceable
2
3@traceable(name="process_document")
4def process_document(doc_text: str):
5    # Step 1: Summarize
6    summary = summarize_chain.invoke({"text": doc_text})
7    
8    # Step 2: Extract entities
9    entities = extract_chain.invoke({"text": doc_text})
10    
11    # Step 3: Classify
12    category = classify_chain.invoke({"text": doc_text})
13    
14    return {
15        "summary": summary,
16        "entities": entities,
17        "category": category
18    }
19
20# Moi call se tao trace voi child spans
21result = process_document("Long document text...")

1from langsmith import Client
2from langsmith.evaluation import evaluate
3
4client = Client()
5
6# Create dataset
7dataset = client.create_dataset("qa-test")
8
9# Add examples
10client.create_examples(
11    inputs=[
12        {"question": "AI la gi?"},
13        {"question": "LangChain la gi?"},
14    ],
15    outputs=[
16        {"answer": "AI la tri tue nhan tao"},
17        {"answer": "LangChain la framework cho LLM apps"},
18    ],
19    dataset_id=dataset.id
20)
21
22# Define evaluator
23def check_relevance(run, example):
24    prediction = run.outputs["output"]
25    reference = example.outputs["answer"]
26    # Simple check
27    return {"score": 1 if len(prediction) > 10 else 0}
28
29# Run evaluation
30results = evaluate(
31    lambda inputs: chain.invoke(inputs),
32    data=dataset.name,
33    evaluators=[check_relevance]
34)

1import wandb
2
3# Initialize run
4wandb.init(
5    project="ai-deployment",
6    config={
7        "model": "gpt-4o-mini",
8        "temperature": 0.7,
9        "max_tokens": 1000
10    }
11)
12
13# Log metrics
14wandb.log({
15    "tokens_used": 150,
16    "latency": 0.45,
17    "cost": 0.001,
18    "quality_score": 8.5
19})
20
21# Finish
22wandb.finish()

1import wandb
2
3# Log results as table
4table = wandb.Table(columns=["input", "output", "tokens", "latency", "score"])
5
6for result in results:
7    table.add_data(
8        result["input"],
9        result["output"],
10        result["tokens"],
11        result["latency"],
12        result["score"]
13    )
14
15wandb.log({"results": table})

1from langchain_core.callbacks import BaseCallbackHandler
2
3class CombinedMonitor(BaseCallbackHandler):
4    """Log to both LangSmith (auto) and W&B"""
5    
6    def __init__(self):
7        self.start_time = None
8    
9    def on_llm_start(self, serialized, prompts, **kwargs):
10        import time
11        self.start_time = time.time()
12    
13    def on_llm_end(self, response, **kwargs):
14        import time
15        latency = time.time() - self.start_time
16        
17        usage = {}
18        if response.llm_output:
19            usage = response.llm_output.get("token_usage", {})
20        
21        # Log to W&B
22        wandb.log({
23            "tokens": usage.get("total_tokens", 0),
24            "latency": latency,
25            "prompt_tokens": usage.get("prompt_tokens", 0),
26            "completion_tokens": usage.get("completion_tokens", 0)
27        })