TCS LangChain Interview Questions 2026

LangChain is an open-source framework for building applications powered by LLMs. It solves the problem of connecting LLMs to external data, tools, and memory. Without LangChain, you write boilerplate for every: prompt formatting, chain-of-thought parsing, tool calling, memory management, and streaming. LangChain provides standardized abstractions so you can swap models (GPT-4 → Gemini) without rewriting business logic.

Chains follow a fixed, predetermined sequence of steps — ideal for workflows where the path is known (query → embed → retrieve → generate). Agents are dynamic: the LLM decides at runtime which tools to call and in what order. Use chains for production stability and cost predictability. Use agents for open-ended tasks requiring reasoning. In production, most teams start with chains and only use agents where dynamic tool selection is genuinely needed.

LangChain Memory stores conversation context to enable multi-turn conversations. Types: (1) ConversationBufferMemory — stores all messages (expensive for long conversations), (2) ConversationSummaryMemory — summarizes old messages with the LLM, (3) ConversationBufferWindowMemory — keeps last N messages, (4) VectorStoreRetrieverMemory — retrieves relevant past messages by semantic similarity. For production, use external memory (Redis/PostgreSQL) instead of in-memory solutions that reset on restart.

LCEL is LangChain's declarative way to compose chains using the pipe operator (|). Example: chain = prompt | llm | output_parser. LCEL gives you: automatic async support, streaming by default, built-in retry logic, parallel execution (RunnableParallel), and easy debugging. Every component is a Runnable with .invoke(), .stream(), and .batch() methods. LCEL is now the recommended way to build all LangChain applications.

Production RAG pipeline: (1) Document loading with UnstructuredLoader, (2) Chunking with RecursiveCharacterTextSplitter (chunk_size=1000, overlap=200), (3) Embedding with OpenAIEmbeddings, (4) Storage in Pinecone/Chroma vectorstore, (5) Retrieval with similarity search (top-k=5), (6) Reranking with Cohere Reranker, (7) Answer generation with ChatOpenAI + prompt template. Monitor retrieval quality with RAGAS evaluation framework.

Tools are functions that agents can call to interact with the world. Define with @tool decorator: @tool def search_web(query: str) -> str. Tools need: name (for LLM to identify), description (for LLM to know when to use), and a callable function. Common tools: web search (Tavily), code execution (Python REPL), database queries, API calls, file operations. In production, validate tool inputs, add timeouts, and handle errors gracefully.

Use LangSmith Hub for centralized prompt storage and versioning. Alternatively, store prompts in a database or YAML files with version tags. Key practices: (1) Never hardcode prompts in code — externalize them, (2) A/B test prompts with different versions, (3) Track which prompt version produced each response, (4) Use PromptTemplate with input_variables for reusable prompts. LangSmith provides prompt testing and comparison tools.

LangGraph is a library built on LangChain for building stateful multi-agent workflows as graphs. Use LangGraph when: (1) You need complex branching logic and loops, (2) Multiple agents collaborate with shared state, (3) You need human-in-the-loop review steps, (4) Your workflow has conditional paths based on intermediate results. For simple single-agent tasks, standard LangChain agents are sufficient.

Use LangSmith for production tracing — it records every LLM call, tool execution, and chain step with timing and cost data. Enable with LANGCHAIN_TRACING_V2=true env variable. For local debugging: set verbose=True on any chain, use callbacks (StdOutCallbackHandler), or use the @traceable decorator. LangSmith lets you replay failed requests, compare runs, and set up automated evaluation pipelines.

Cost reduction strategies: (1) Caching — use LangChain's built-in cache (SQLiteCache or RedisCache) to avoid duplicate API calls, (2) Model routing — use GPT-3.5 for simple queries, GPT-4 only for complex reasoning, (3) Context compression — use ContextualCompressionRetriever to reduce tokens sent to LLM, (4) Batch processing — use .batch() instead of individual .invoke() calls, (5) Token counting — set max_tokens limits and monitor with LangSmith.