Skip to main content

Langchain

npm install @trpc-chat-agent/langchain

Setting up LangChain is fairly straightforward

import { langchainBackend, asLangChainMessagesArray, models } from '@trpc-chat-agent/langchain';
import { initAgents } from '@trpc-chat-agent/core';
import { ChatOpenAI } from 'langchain/chatmodels/openai';

export const ai = initAgents
  .context<typeof createContext>()
  .backend(langchainBackend) // LangChain adapter
  .create();

// Now initialize your agent with these parameters:
const agent = ai.agent({
  // The tools for your agent
  tools: [myTool1, myTool2],

  // LLM, from LangChain. Must extend BaseChatModel
  // Can either be:
  // - a single model
  // - a function to select what model to use based on extra args
  // Here's a single model:
  llm: models.openai({ model: 'gpt-4o' }),
  // Here's a map of named models:
  llm: ({ extraArgs }) => models.openai({ model: extraArgs.myChosenModelName }),

  // (optional) The maximum update rate for client args and progress indicators
  // defaults to 100ms
  debounceMs: 100,

  // (optional) Implement a custom message transformer. `ctx` is the tRPC context.
  // Uses the implementation shown here if you don't provide one
  transformMessages: async ({ conversation, path, ctx }) => {
    return asLangChainMessagesArray(conversation, path);
  },

  // (optional) System message to append at the beginning. Gets
  // appended after transformMessages runs
  systemMessage: 'You are a helpful assistant',

  // (optional) LangChain callbacks, e.g. for custom tracing
  langchainCallbacks: [customTracingCallback],
});

When defining tool calls, you have access to LangChain's extra function parameters:

  • Callbacks
  • Runnable config

You can treat the run call the same as the callbacks in LangChain's DynamicStructuredTool.

const expertModel = new ChatOpenAI({ model: 'gpt-6-pro' });

const myTool = ai.tool({
  name: 'callAgent',
  description: 'Ask an AI agent for help',
  schema: z.object({
    question: z.string(),
  }),

  // When calling `run`, the tool has access to LangChain's:
  // - Callbacks
  // - Runnable config
  run: async ({ input, /* ... */ }, langchainCallbacks, runnableConfig) => {
    // Invoke other LangChain APIs
    const chatResult = await expertModel.call({
      messages: [new UserMessage(input.question)],
    });
    const result = chatResult.content;

    return {
      response: result,
      clientResult: null,
    };
  },
});