LangChain & LangGraph Reference
Glossary, diagram atlas, and the APIs you will actually touch when the demo graduates into a system.
Version posture
| Stack | Pinned in chapters | Why it matters |
|---|---|---|
| LangChain | 0.3.x | Runnable-first APIs, fewer legacy chain ghosts. |
| LangGraph | 0.2.x | StateGraph, interrupts, checkpointers, and the good stuff. |
| Python | 3.11+ | Modern async ergonomics without pretending 3.8 is fine forever. |
Diagram atlas
- lc-langchain-ecosystem-map — What LangChain Actually Is — And why you stopped calling OpenAI directly.
- lc-runnable-protocol-shape — What LangChain Actually Is — And why you stopped calling OpenAI directly.
- lc-prompt-template-render — Models, Prompts, Parsers — The holy trinity, and the part where we stop pretending regex works.
- lc-structured-output-paths — Models, Prompts, Parsers — The holy trinity, and the part where we stop pretending regex works.
- lc-lcel-composition — LCEL — The pipe that actually earns its keep.
- lc-lcel-parallel-branch — LCEL — The pipe that actually earns its keep.
- lc-rag-pipeline-naive — Retrieval Fundamentals — Why your first RAG app was a glorified substring search.
- lc-chunk-overlap-anatomy — Retrieval Fundamentals — Why your first RAG app was a glorified substring search.
- lc-rag-hybrid-with-reranker — Advanced RAG — Hybrid search, re-ranking, and the parent-document trick the demos skipped.
- lc-parent-doc-retriever — Advanced RAG — Hybrid search, re-ranking, and the parent-document trick the demos skipped.
- lc-multi-query-fanout — Advanced RAG — Hybrid search, re-ranking, and the parent-document trick the demos skipped.
- lc-tool-call-roundtrip — Tools and Function Calling — Stop pretending the model picks for you.
- lc-parallel-tool-execution — Tools and Function Calling — Stop pretending the model picks for you.
- lc-chain-vs-graph — LangGraph Fundamentals — Graphs, because AgentExecutor finally got the memo.
- lc-stategraph-anatomy — LangGraph Fundamentals — Graphs, because AgentExecutor finally got the memo.
- lc-state-reducer-flow — Stateful Agents — Messages, reducers, and memory that does not lie to itself.
- lc-message-trim-strategies — Stateful Agents — Messages, reducers, and memory that does not lie to itself.
- lc-supervisor-pattern — Multi-Agent Patterns — Supervisor, swarm, and knowing when one agent was fine, actually.
- lc-swarm-handoff — Multi-Agent Patterns — Supervisor, swarm, and knowing when one agent was fine, actually.
- lc-hierarchical-teams — Multi-Agent Patterns — Supervisor, swarm, and knowing when one agent was fine, actually.
- lc-interrupt-resume-cycle — Human-in-the-Loop — Interrupts, approvals, and the time-travel debugger you did not know you had.
- lc-time-travel-fork — Human-in-the-Loop — Interrupts, approvals, and the time-travel debugger you did not know you had.
- lc-checkpointer-thread-tree — Persistence and Checkpointers — How to survive a pod restart without a customer noticing.
- lc-checkpoint-backends-tradeoffs — Persistence and Checkpointers — How to survive a pod restart without a customer noticing.
- lc-astream-events-lifecycle — Streaming, Async, and Concurrency — Sync code in 2026 is a personality choice.
- lc-parallel-vs-sequential-branches — Streaming, Async, and Concurrency — Sync code in 2026 is a personality choice.
- lc-injection-vectors — Security and Guardrails — Your agent is an untrusted browser with a wallet.
- lc-defense-in-depth-layers — Security and Guardrails — Your agent is an untrusted browser with a wallet.
- lc-cost-flow-per-node — Cost and Token Budgets — How to not wake up to a five-figure bill you cannot explain in Slack.
- lc-model-router-decision — Cost and Token Budgets — How to not wake up to a five-figure bill you cannot explain in Slack.
- lc-trace-tree-anatomy — Observability with LangSmith — Traces that save you at 3 AM, plus the self-hosting argument.
- lc-langsmith-self-host-vs-cloud — Observability with LangSmith — Traces that save you at 3 AM, plus the self-hosting argument.
- lc-eval-loop — Evaluations — LLM-as-judge without lying to yourself.
- lc-judge-calibration — Evaluations — LLM-as-judge without lying to yourself.
- lc-langserve-route-surface — Deployment — LangServe, FastAPI, and why your serverless function is about to time out.
- lc-deployment-topologies — Deployment — LangServe, FastAPI, and why your serverless function is about to time out.
- lc-capstone-architecture — Capstone — Ship a production multi-agent research assistant, or admit you are still doing demos.
- lc-capstone-data-flow — Capstone — Ship a production multi-agent research assistant, or admit you are still doing demos.
- lc-capstone-trace-anatomy — Capstone — Ship a production multi-agent research assistant, or admit you are still doing demos.
- lc-playbook-pre-flight-checklist — LangChain & LangGraph Playbook
- lc-playbook-on-call-decision — LangChain & LangGraph Playbook
Glossary
Runnable
The core LangChain interface for invokable units that support sync, async, batch, stream, config, and composition.
LCEL
LangChain Expression Language: the pipe-friendly composition layer for building runnable graphs without inventing another framework in a meeting.
StateGraph
A LangGraph graph definition whose nodes read and return typed state updates.
Checkpointer
Durable storage for graph state snapshots, resumes, interrupts, and time-travel debugging.
Thread
A logical conversation or run identity used by checkpointers to resume the right graph state.
Reducer
A function that merges node updates into graph state, usually where message history gets less cursed.
Tool
A callable capability exposed to a model or graph node with a name, description, and schema.
bind_tools
The ChatModel method that attaches tool schemas so the model can request calls instead of hallucinating APIs.
Retriever
A component that returns relevant documents for a query from a corpus or search backend.
Re-ranker
A second-stage scorer that reorders retrieved candidates by relevance before the model sees them.
Parent-Document Retriever
A retrieval pattern that searches smaller chunks but returns larger parent documents for usable context.
Hybrid Search
Retrieval that blends lexical matching with dense vectors, because embeddings alone are not a personality.
Interrupt
A LangGraph pause point that asks an external actor to approve, edit, or provide data before continuing.
Time Travel
Checkpoint-based replay or fork of a graph run from an earlier state.
Supervisor
A routing agent that decides which worker or node handles the next step.
Swarm
A multi-agent handoff pattern where control moves between peers instead of one boss router.
LangSmith Run
A traced unit of execution with inputs, outputs, metadata, child runs, and enough evidence to debug the weird thing.
Evaluator
A function or model-backed judge that scores outputs against expected behavior.
LLM-as-Judge
An evaluation pattern where a model grades model output, useful when calibrated and comedy when not.
Prompt Hub
A registry for versioning prompts instead of hiding them in heroic Python strings.
LangServe Route
An HTTP endpoint generated around a runnable for invoke, batch, stream, and playground use.
astream_events
The event stream API for observing runnable progress, tokens, node boundaries, and custom events.
Pydantic Args
Typed tool argument schemas that prevent the model from sending nonsense with confidence.
with_structured_output
A model wrapper that coerces responses into a typed schema instead of asking JSON politely.
MessagesState
A LangGraph state helper for chat histories with message-aware merging.