NEW LangChain Expression Language!!

LangChain’s new Expression Language is a more declarative way to build LLM “chains,” making the flow of data through prompts, models, tools, and retrievers much easier to read and debug. After months of criticism aimed at confusing documentation and an overly complex API, the update focuses on clarity: developers can now see how inputs move step-by-step under the hood, using a syntax built around piping components together.

At the core, LangChain treats applications as compositions of prompts and LLM calls, with optional steps that incorporate other capabilities—like vector search for context, function calling, streaming/batching/async execution, and tool usage. The Expression Language introduces a “new syntax” that defines chains as connected components. In the simplest form, a prompt is piped into a model, producing an output. That same chain can then be executed in different modes using explicit methods: invoke for standard runs, batch for batch processing, and stream for streaming outputs. The result is less guesswork about what each chain does and fewer surprises about execution behavior.

The transcript walks through practical examples that show how the syntax changes day-to-day development. A basic chain can be defined as `prompt | model`, then optionally followed by an output parser to control the returned type—either an AI message object or a plain string. The same structure works across chat models and older text-style models, including Hugging Face models and OpenAI-style models like `text-davinci-003`, with the chain definition staying largely the same while the model changes.

Bindings add another layer of control. One example uses a stop binding to limit generation—such as stopping after the first line so a “three facts” prompt returns only one item. Another binding integrates OpenAI function calling by attaching a function schema to the model. In that setup, the model returns structured JSON-like arguments (e.g., a joke split into required fields like setup and punchline). Output passes then let developers extract either the full structured result or specific keys from it.

Retrievers demonstrate how the language handles data reuse inside a chain. When a question must be used both to fetch vector-store context and again to ask the model, the syntax uses pass-through mechanisms (like item getters) to route the same input to multiple steps without transformation. A Chroma vector store with OpenAI embeddings is used to retrieve context for questions like “who is James Bond,” then feed that context and the original question into a prompt. The example extends to multilingual answering by passing a language parameter and selecting it via dictionary-based getters.

Tools and runnable functions round out the picture. A DuckDuckGo example shows a non-agent workflow: a prompt rewrites a query for search, the search runs, and the results can be fed forward. Another example uses arbitrary runnable functions to compute values (like string length and arithmetic composition) before sending the final prompt to the model.

Overall, the Expression Language replaces much of the older chain-building approach with a clearer, pipe-based syntax. Agents aren’t the focus here, but the update positions future agent work to reuse the same declarative style, while delivering immediate benefits for summarization, chat workflows, and tool-augmented pipelines.