LangChain

LangChain is the most popular LLM application development framework, covering the full pipeline from prototype to production. Core modules: langchain-core (abstractions) / langchain (integrations) / langgraph (graph orchestration) / langsmith (observability).

Core concepts

• Model I/O: unified wrappers for LLMs, chat models, embeddings.
• Prompt Templates: templated prompts with variables, few-shot support.
• Output Parsers: parse LLM output into structured data.
• Chains: compose multiple steps into a pipeline.
• Agents: LLM-driven decision loops + tool use.
• Retrievers: retrieval abstraction layer, unified across vector stores.

LangGraph: state-machine agents

LangGraph models an Agent as a state graph (StateGraph):

• Nodes: each step (LLM call, tool execution, human input).
• Edges: transition conditions (routing, loops, conditionals).
• State: shared state object across nodes (TypedDict).

Good for complex flows: multi-agent, human-in-the-loop, recoverable execution.

LangSmith: observability

• Trace every LLM call, tool execution, token usage, latency.
• Online evaluation datasets.
• Prompt version management + A/B testing.
• Production monitoring and alerts.

Good for

• Agent / multi-step tool use: LangGraph first.
• Complex RAG / retrieval augmentation: LangChain + LlamaIndex mix.
• Production observability: LangSmith essential.

Caveats

• Frequent API iteration: many breaking changes between versions.
• Sometimes over-abstracted: simple needs take longer paths.
• Steep learning curve: Chain / Agent / Graph / Expression Language are many concepts.