feat(core): Instrument LangChain AI

[RulaKhaled]

This PR adds automatic instrumentation for LangChain chat clients in Node SDK, we cover most used providers mentioned in https://python.langchain.com/docs/integrations/chat/.

What's added?

TLDR; a LangChain Callback Handler that:

• Creates a stateful callback handler that tracks LangChain lifecycle events
• Handles LLM/Chat Model events (start, end, error, streaming)
• Handles Chain events (start, end, error)
• Handles Tool events (start, end, error)
• Extracts and normalizes request/respo

How it works?

1. Module Patching: When a LangChain provider package is loaded (e.g., @langchain/anthropic), the instrumentation:

   • Finds the chat model class (e.g., ChatAnthropic)
   • Wraps the invoke, stream, and batch methods on the prototype
   • Uses a Proxy to intercept method calls

2. Callback Injection: When a LangChain method is called:

   • The wrapper intercepts the call
   • Augments the options.callbacks array with Sentry's callback handler
   • Calls the original method with the augmented callbacks

The integration is enabled by default when you initialize Sentry in Node.js:

import * as Sentry from '@sentry/node';
import { ChatAnthropic } from '@langchain/anthropic';

Sentry.init({
  dsn: 'your-dsn',
  tracesSampleRate: 1.0,
  sendDefaultPii: true, // Enable to record inputs/outputs
});

// LangChain calls are automatically instrumented
const model = new ChatAnthropic({
  model: 'claude-3-5-sonnet-20241022',
});

await model.invoke('What is the capital of France?');

You can configure what data is recorded:

Sentry.init({
  integrations: [
    Sentry.langChainIntegration({
      recordInputs: true,  // Record prompts/messages
      recordOutputs: true, // Record responses
    })
  ],
});

Note: We need to disable integrations for AI providers that LangChain use to avoid duplicate spans, this will be handled in a follow up PR.

[github-actions]

size-limit report 📦

Path	Size	% Change	Change
@sentry/browser	24.63 kB	-	-
@sentry/browser - with treeshaking flags	23.11 kB	-	-
@sentry/browser (incl. Tracing)	40.97 kB	-	-
@sentry/browser (incl. Tracing, Profiling)	45.26 kB	-	-
@sentry/browser (incl. Tracing, Replay)	79.29 kB	-	-
@sentry/browser (incl. Tracing, Replay) - with treeshaking flags	68.97 kB	-	-
@sentry/browser (incl. Tracing, Replay with Canvas)	83.99 kB	-	-
@sentry/browser (incl. Tracing, Replay, Feedback)	96.16 kB	-	-
@sentry/browser (incl. Feedback)	41.3 kB	-	-
@sentry/browser (incl. sendFeedback)	29.29 kB	-	-
@sentry/browser (incl. FeedbackAsync)	34.22 kB	-	-
@sentry/react	26.31 kB	-	-
@sentry/react (incl. Tracing)	42.97 kB	-	-
@sentry/vue	29.11 kB	-	-
@sentry/vue (incl. Tracing)	42.75 kB	-	-
@sentry/svelte	24.64 kB	-	-
CDN Bundle	26.9 kB	-	-
CDN Bundle (incl. Tracing)	41.62 kB	-	-
CDN Bundle (incl. Tracing, Replay)	77.87 kB	-	-
CDN Bundle (incl. Tracing, Replay, Feedback)	83.35 kB	-	-
CDN Bundle - uncompressed	78.86 kB	-	-
CDN Bundle (incl. Tracing) - uncompressed	123.44 kB	-	-
CDN Bundle (incl. Tracing, Replay) - uncompressed	238.48 kB	-	-
CDN Bundle (incl. Tracing, Replay, Feedback) - uncompressed	251.24 kB	-	-
@sentry/nextjs (client)	45.11 kB	-	-
@sentry/sveltekit (client)	41.4 kB	-	-
@sentry/node-core	50.75 kB	-	-
@sentry/node	156.91 kB	+1.28%	+1.98 kB 🔺
@sentry/node - without tracing	92.61 kB	+0.01%	+1 B 🔺
@sentry/aws-serverless	106.33 kB	+0.01%	+1 B 🔺

View base workflow run

[github-actions]

node-overhead report 🧳

Note: This is a synthetic benchmark with a minimal express app and does not necessarily reflect the real-world performance impact in an application.

Scenario	Requests/s	% of Baseline	Prev. Requests/s	Change %
GET Baseline	9,190	-	11,704	-21%
GET With Sentry	1,354	15%	1,613	-16%
GET With Sentry (error only)	6,016	65%	7,913	-24%
POST Baseline	1,187	-	1,216	-2%
POST With Sentry	508	43%	564	-10%
POST With Sentry (error only)	1,050	88%	1,068	-2%
MYSQL Baseline	3,248	-	4,148	-22%
MYSQL With Sentry	352	11%	593	-41%
MYSQL With Sentry (error only)	2,662	82%	3,405	-22%

View base workflow run

[andreiborza]

LGTM, nice work!