Instrumenting with LangSmith

[eric-gardyn]

Some feedback:
I am looking into LangSmith as a tool to monitor and audit the queries.
One of the steps is to retrieve the result from openAiClient.getChatCompletions, and send it to LangSmith.

In my project, I have now copied all the mongodb-server package files so I can customize/debug as needed.
So I added code into the answerQuestionAwaited method to interact with LangSmith.

Wondering if this could be done another way (so I could eventually just use the mongodb-server package "as is")?

[nlarew]

Hey Eric!

Interesting use case. Right now I think the best approach in our framework is more or less what you've done, i.e. create a custom ChatLlm interface in your config instead of using our makeOpenAiChatLlm() constructor. We can consider ways to extend our implementation so that you don't have to write a totally custom implementation.

We haven't worked with LangSmith so I'm not super certain of how LangChain interacts with it. If there's a good integration there, you might also consider using the makeLangchainChatLlm() constructor with a LangChain OpenAI ChatModel instance. Do you think that would be useful in this case?

[nlarew]

@eric-gardyn if you can would you mind sharing the changes you made to answerQuestionAwaited() to interact with LangSmith? I'd like to get a feel for what this type of integration looks like so that we could make it easier in future.

[eric-gardyn]

sure, it's quick-and-dirty, but straight-forward, just so I could get it up and running

import { RunTree } from 'langsmith'
(...)
   async answerQuestionAwaited(
      { messages, toolCallOptions }: LlmAnswerQuestionParams,
      question: string
    ) {
      const pipeline = new RunTree({
        name: 'Chat Pipeline',
        run_type: 'chain',
        inputs: { question },
      })

      // Create a child run
      const childRun = await pipeline.createChild({
        name: 'OpenAI Call',
        run_type: 'llm',
        inputs: { messages },
      })

      const chatCompletion = await openAiClient.getChatCompletions(deployment, messages, {
        ...openAiLmmConfigOptions,
        ...(toolCallOptions ? { functionCall: toolCallOptions } : {}),
        functions: tools?.map(tool => {
          return tool.definition
        }),
      })

      const {
        choices: [choice],
      } = chatCompletion
      const { message } = choice
      if (!message) {
        throw new Error('No message returned from OpenAI')
      }

      // End the runs and log them
      childRun.end(chatCompletion)
      await childRun.postRun()

      pipeline.end({ outputs: { answer: message.content } })
      await pipeline.postRun()

      return message as ChatRequestAssistantMessage
    },

and I changed awaitGenerateResponse to call 'answerQuestionAwaited' with added parameter 'request?.body?.message'

and

export interface ChatLlm {
  answerQuestionStream(params: LlmAnswerQuestionParams): Promise<OpenAiStreamingResponse>
  answerQuestionAwaited(
    params: LlmAnswerQuestionParams,
    question?: string
  ): Promise<OpenAiAwaitedResponse>
  callTool?(params: LlmCallToolParams): Promise<CallToolResponse>
}

there might be an easier/better way