How to create a custom LLM class
This guide assumes familiarity with the following concepts:
This notebook goes over how to create a custom LLM wrapper, in case you want to use your own LLM or a different wrapper than one that is directly supported in LangChain.
There are a few required things that a custom LLM needs to implement
after extending the LLM
class:
- A _callmethod that takes in a string and call options (which includes things likestopsequences), and returns a string.
- A _llmTypemethod that returns a string. Used for logging purposes only.
You can also implement the following optional method:
- A _streamResponseChunksmethod that returns anAsyncIteratorand yieldsGenerationChunks. This allows the LLM to support streaming outputs.
Letβs implement a very simple custom LLM that just echoes back the first
n characters of the input.
import { LLM, type BaseLLMParams } from "@langchain/core/language_models/llms";
import type { CallbackManagerForLLMRun } from "@langchain/core/callbacks/manager";
import { GenerationChunk } from "@langchain/core/outputs";
interface CustomLLMInput extends BaseLLMParams {
  n: number;
}
class CustomLLM extends LLM {
  n: number;
  constructor(fields: CustomLLMInput) {
    super(fields);
    this.n = fields.n;
  }
  _llmType() {
    return "custom";
  }
  async _call(
    prompt: string,
    options: this["ParsedCallOptions"],
    runManager: CallbackManagerForLLMRun
  ): Promise<string> {
    // Pass `runManager?.getChild()` when invoking internal runnables to enable tracing
    // await subRunnable.invoke(params, runManager?.getChild());
    return prompt.slice(0, this.n);
  }
  async *_streamResponseChunks(
    prompt: string,
    options: this["ParsedCallOptions"],
    runManager?: CallbackManagerForLLMRun
  ): AsyncGenerator<GenerationChunk> {
    // Pass `runManager?.getChild()` when invoking internal runnables to enable tracing
    // await subRunnable.invoke(params, runManager?.getChild());
    for (const letter of prompt.slice(0, this.n)) {
      yield new GenerationChunk({
        text: letter,
      });
      // Trigger the appropriate callback
      await runManager?.handleLLMNewToken(letter);
    }
  }
}
We can now use this as any other LLM:
const llm = new CustomLLM({ n: 4 });
await llm.invoke("I am an LLM");
I am
And support streaming:
const stream = await llm.stream("I am an LLM");
for await (const chunk of stream) {
  console.log(chunk);
}
I
a
m
If you want to take advantage of LangChainβs callback system for
functionality like token tracking, you can extend the
BaseLLM
class and implement the lower level _generate method. Rather than
taking a single string as input and a single string output, it can take
multiple input strings and map each to multiple string outputs.
Additionally, it returns a Generation output with fields for
additional metadata rather than just a string.
import { CallbackManagerForLLMRun } from "@langchain/core/callbacks/manager";
import { LLMResult } from "@langchain/core/outputs";
import {
  BaseLLM,
  BaseLLMCallOptions,
  BaseLLMParams,
} from "@langchain/core/language_models/llms";
interface AdvancedCustomLLMCallOptions extends BaseLLMCallOptions {}
interface AdvancedCustomLLMParams extends BaseLLMParams {
  n: number;
}
class AdvancedCustomLLM extends BaseLLM<AdvancedCustomLLMCallOptions> {
  n: number;
  constructor(fields: AdvancedCustomLLMParams) {
    super(fields);
    this.n = fields.n;
  }
  _llmType() {
    return "advanced_custom_llm";
  }
  async _generate(
    inputs: string[],
    options: this["ParsedCallOptions"],
    runManager?: CallbackManagerForLLMRun
  ): Promise<LLMResult> {
    const outputs = inputs.map((input) => input.slice(0, this.n));
    // Pass `runManager?.getChild()` when invoking internal runnables to enable tracing
    // await subRunnable.invoke(params, runManager?.getChild());
    // One input could generate multiple outputs.
    const generations = outputs.map((output) => [
      {
        text: output,
        // Optional additional metadata for the generation
        generationInfo: { outputCount: 1 },
      },
    ]);
    const tokenUsage = {
      usedTokens: this.n,
    };
    return {
      generations,
      llmOutput: { tokenUsage },
    };
  }
}
This will pass the additional returned information in callback events and in the `streamEvents method:
const llm = new AdvancedCustomLLM({ n: 4 });
const eventStream = await llm.streamEvents("I am an LLM", {
  version: "v2",
});
for await (const event of eventStream) {
  if (event.event === "on_llm_end") {
    console.log(JSON.stringify(event, null, 2));
  }
}
{
  "event": "on_llm_end",
  "data": {
    "output": {
      "generations": [
        [
          {
            "text": "I am",
            "generationInfo": {
              "outputCount": 1
            }
          }
        ]
      ],
      "llmOutput": {
        "tokenUsage": {
          "usedTokens": 4
        }
      }
    }
  },
  "run_id": "a9ce50e4-f85b-41eb-bcbe-793efc52f9d8",
  "name": "AdvancedCustomLLM",
  "tags": [],
  "metadata": {}
}