Skip to main content
This guide wires LangSight into a LangChain retrieval-augmented generation (RAG) pipeline. By the end you will see every retrieval call, LLM call, and tool call in the LangSight sessions dashboard with latency, status, and cost.

​
What you need

  • LangSight running locally (./scripts/quickstart.sh or docker compose up -d)
  • Python 3.11+
  • An OpenAI API key (or swap for any LangChain-supported LLM)

​
Install

pip install langsight langchain langchain-openai langchain-community

​
1. Create a project and get your API key

  1. Open http://localhost:3003 and log in
  2. Go to Settings → Projects and copy your project ID
  3. Go to Settings → API Keys and create a key
Add to .env: 
LANGSIGHT_URL=http://localhost:8000
LANGSIGHT_API_KEY=ls_your_key_here
LANGSIGHT_PROJECT_ID=your_project_id

​
2. Build a RAG chain with tracing

import os
from dotenv import load_dotenv

from langchain_openai import ChatOpenAI, OpenAIEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.schema.runnable import RunnablePassthrough
from langchain.prompts import ChatPromptTemplate

import langsight
from langsight.sdk import LangSightClient
from langsight.integrations.langchain import LangSightLangChainCallback

load_dotenv()

# --- LangSight setup ---
client = LangSightClient(
    url=os.environ["LANGSIGHT_URL"],
    api_key=os.environ["LANGSIGHT_API_KEY"],
    project_id=os.environ["LANGSIGHT_PROJECT_ID"],
)
callback = LangSightLangChainCallback(
    client=client,
    agent_name="rag-agent",
)

# --- Build a minimal FAISS vector store from sample docs ---
docs = [
    "LangSight monitors AI agent tool calls in production.",
    "The circuit breaker disables a failing tool after 5 consecutive errors.",
    "Loop detection fires when the same tool is called 3 times with identical arguments.",
    "Cost attribution shows which MCP server is consuming the most budget.",
]
splitter = RecursiveCharacterTextSplitter(chunk_size=100, chunk_overlap=0)
chunks = splitter.create_documents(docs)
vectorstore = FAISS.from_documents(chunks, OpenAIEmbeddings())
retriever = vectorstore.as_retriever(search_kwargs={"k": 2})

# --- Build the RAG chain ---
prompt = ChatPromptTemplate.from_template(
    "Answer based on this context:\n\n{context}\n\nQuestion: {question}"
)
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

def format_docs(docs):
    return "\n\n".join(d.page_content for d in docs)

rag_chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
)

# --- Run with tracing ---
with langsight.session(agent_name="rag-agent") as session_id:
    callback.session_id = session_id
    response = rag_chain.invoke(
        "How does LangSight detect agent loops?",
        config={"callbacks": [callback]},
    )
    print(response.content)

​
3. View the trace

langsight sessions

Session          Agent        Calls   Failed   Duration   Cost
sess-abc123      rag-agent    3       0        1.2s       $0.0004

langsight sessions --id sess-abc123

Trace: sess-abc123  (rag-agent)
├── rag-agent (agent)
│   ├── VectorStoreRetriever    340ms   success
│   └── ChatOpenAI              820ms   success
Or open http://localhost:3003 → Sessions → click the session row → Trace tab for the full nested call tree.

​
What gets traced

SpanWhat it captures
Retriever callTool name, latency, status
LLM callModel, input/output tokens, cost, latency
Agent spanWraps all calls in the session
The callback captures the question as the session input and the LLM response as the session output — both visible in the session detail view.

​
Using a different LLM

Swap ChatOpenAI for any LangChain-compatible model. The callback is LLM-agnostic — it traces at the LangChain callback layer, not the model SDK layer. 
from langchain_anthropic import ChatAnthropic
llm = ChatAnthropic(model="claude-3-5-haiku-20241022")

​
Loop detection and budget guardrails

If the retriever is called in a loop (e.g. a retry chain that keeps fetching), LangSight can stop the session automatically. Add to your client: 
client = LangSightClient(
    url=os.environ["LANGSIGHT_URL"],
    api_key=os.environ["LANGSIGHT_API_KEY"],
    project_id=os.environ["LANGSIGHT_PROJECT_ID"],
    loop_detection=True,   # stop if same tool called 3x with same args
    max_cost_usd=0.10,     # hard budget cap per session
)

​
Next steps