Migrating from LangChain.js

Planned, not yet detailed

This page is a placeholder. A full migration guide ships with v0.2 alongside the LangChain interop adapter.

Why these are complementary, not competing

llm-ports is a utility, not a framework. LangChain.js is a framework with chains, agents, retrieval, memory, and prompt templates as first-class concepts. They solve different problems.

If you're using LangChain.js for chains and retrieval, you don't need to leave it. What you can do today is wrap LangChain's underlying LLM calls with an llm-ports port to get cost gating, fallback chains, and provider routing on top — without abandoning chains, retrievers, or vector stores.

The pragmatic v0.1 pattern

LangChain's ChatAnthropic / ChatOpenAI / etc. take provider config in their constructors. You can't directly swap them for an LLMPort. But for the LLM-call segments of your pipeline, you can:

1. Compute the prompt with LangChain (templates, retrieval, chains).
2. Hand it off to llm-ports for the actual model call.
3. Pass the result back to LangChain for downstream chain steps.

import { llm } from "./llm-ports-setup.js";    // your llm-ports port

// ... LangChain chain that produces a final prompt string ...
const promptText = await myChain.format({ ... });

// LLM call goes through llm-ports for cost gating + fallback
const result = await llm.generateText({
  taskType: "draft",
  prompt: promptText,
});

// Continue downstream with LangChain (output parsing, memory, etc.)
await myChain.parseOutput(result.text);

You lose some of LangChain's abstractions for the LLM call itself (the chain doesn't know about llm-ports), but you gain the production primitives llm-ports ships.

What v0.2 adds

@llm-ports/adapter-langchain will provide a BaseChatModel subclass that you can drop into LangChain chains as if it were ChatAnthropic / ChatOpenAI. Underneath, it routes to llm-ports. That gets you:

• LangChain's chains, retrievers, vector stores — unchanged
• llm-ports's cost gating, fallback chains, capability factories — applied to every LLM call

Until v0.2 ships, the manual pattern above is the bridge.

Reading next

• Migrating from a direct SDK →
• llm-ports philosophy →