@dltech/nestjs-langchain

NestJS-friendly utilities for building on top of LangChain:

• Ai* decorators + classToToolCall/ClassToolOutputParser — describe a tool/function call's arguments as a plain TypeScript class and get a JSON Schema FunctionDefinition plus a streaming output parser that validates and instantiates the result, instead of hand-writing JSON Schema.
• llmFallback — chain multiple Runnables as ordered fallbacks.
• streamToRxjs — adapt a LangChain IterableReadableStream to an RxJS Observable.
• ChainProvider — a typed NestJS FactoryProvider shape for chains built from Runnables.

Install

npm install @dltech/nestjs-langchain

Peer dependencies you're expected to already have in your app: @langchain/core, langchain, @nestjs/common, class-transformer, class-validator, reflect-metadata, rxjs, zod.

Usage

Describing a tool call as a class

import { AiToolCall, AiString, AiNumber, AiOptional, ClassToolOutputParser } from '@dltech/nestjs-langchain';

@AiToolCall({ description: 'Book a restaurant reservation' })
class BookReservationArgs {
  @AiString('Name of the restaurant')
  restaurant: string;

  @AiNumber('Party size')
  partySize: number;

  @AiString('Requested time, ISO 8601')
  @AiOptional
  time?: string;
}

const parser = new ClassToolOutputParser(BookReservationArgs);
const { tool_choice, tools } = parser.bindLLM();

const modelWithTools = model.bind({ tools, tool_choice });
const result = await modelWithTools.pipe(parser).invoke(messages);
// result is a BookReservationArgs, validated with class-validator

Falling back between models

import { llmFallback } from '@dltech/nestjs-langchain';

const chain = llmFallback(primaryModel, secondaryModel, tertiaryModel);

Streaming a LangChain stream into RxJS

import { streamToRxjs } from '@dltech/nestjs-langchain';

const stream = await chain.stream(input);
streamToRxjs(stream).subscribe((chunk) => sendToClient(chunk));