LangChain

LangChain ships MRKLChain, an implementation named after the MRKL (Modular Reasoning, Knowledge and Language) paper: an LLM dispatcher loops over tool definitions and, at each step, picks one action…

RAG is documented as 'a simple agent with a tool that retrieves information' — the model decides when to call the retrieval tool, mirroring the agentic-RAG shape.

CrossEncoderReranker wraps a HuggingFace cross-encoder and plugs into ContextualCompressionRetriever.

RunnableWithFallbacks (or .with_fallbacks()) is a documented core abstraction; fallbacks are tried in order until one succeeds.

First-class MCP integration via langchain-mcp-adapters and MultiServerMCPClient.

LangChain ships MultiQueryRetriever as the canonical Query Rewriting implementation: an LLM generates N variants of the user query, retrieval runs per variant, results are unioned and deduplicated. D…

create_agent IS the prebuilt ReAct constructor.

Threads scope short-term memory; storing state in graph state maintains separation between threads.

response_format with ProviderStrategy (native) or ToolStrategy (tool-call emulation). As of LangChain 1.0, passing a schema defaults to ProviderStrategy and falls back to ToolStrategy.

Tools are the primary extension point; @tool decorator + model.bind_tools(...).

create_self_ask_with_search_agent builds an agent that breaks a question into explicit follow-up sub-questions, answers each (via an Intermediate Answer search tool), then composes the final answer.

EnsembleRetriever (langchain_classic) combines multiple retrievers via weighted Reciprocal Rank Fusion; BM25Retriever (langchain-community) provides keyword search alongside vector retrieval.

Documented in LangChain Inc.'s blog rather than the docs proper; concrete implementations live in langgraph examples. Downgraded from first-class.

Agents stop on final-answer-or-iteration-limit but no current docs page names the iteration-limit parameter; recursion_limit is a LangGraph concept. Downgraded from first-class.

LangChain ships RedisSemanticCache, an LLM cache that matches prompts by vector similarity rather than exact string equality and returns the stored response on a near-duplicate hit.

LLM-as-judge evaluators send one model's output to a separate judge LLM, and the judging model is configurable (PROVIDER:MODEL string or a different chat model instance), decorrelating critique from…

Human-in-the-Loop middleware uses LangGraph's interrupt + checkpointer underneath; halts execution and saves state. Limited because the durability comes from LangGraph, not LangChain proper.

Not a dedicated CoT API; reasoning is exposed as 'reasoning' content blocks during streaming. The ReAct loop interleaves reasoning steps with tool calls.

Citation is a typed content-block annotation; streaming surfaces content blocks but there is no dedicated 'citation streaming' mode named in docs.

Short-term-memory docs name the problem and offer SummarizationMiddleware + token trimming as the answer.

Agents transfer control via tool calls (Command(goto=...) is a LangGraph primitive re-exported by LangChain).

Reflection agents are blog-only; no dedicated docs page in current LangChain.

Skills live primarily in Deep Agents (a higher-level harness built on LangChain), not core LangChain.

langchain_experimental ships a ToTChain that searches a tree of thoughts: a thought-generation strategy proposes children, a ToTChecker classifies each thought VALID_INTERMEDIATE/VALID_FINAL/INVALID,…

No canonical LangChain docs page for Anthropic's Contextual Retrieval technique; usable only as a recipe via custom embedding-time chunking. Downgraded from limited to via-integration. No verbatim up…

Closest core-LangChain concept is the Router (LLM picks among agents and invokes in parallel). The supervisor pattern is documented as a multi-agent architecture with a central supervisor coordinatin…