convex-langfuse

Send LLM traces, generations, and scores from your Convex app to Langfuse. Reactive local queries, OpenTelemetry-native ingestion.

const langfuse = new Langfuse(components.convexLangfuse, {
  publicKey: process.env.LANGFUSE_PUBLIC_KEY!,
  secretKey: process.env.LANGFUSE_SECRET_KEY!,
});

// Log a completed LLM call
const { traceId } = await langfuse.logGeneration(ctx, {
  name: "chat-completion",
  userId: "user_123",
  model: "gpt-5",
  input: prompt,
  output: completion,
  inputTokens: 120,
  outputTokens: 340,
});

// Attach a user rating or eval score to it
await langfuse.recordScore(ctx, { traceId, name: "user-rating", value: 1 });

What this does

Most Langfuse integrations assume a long-lived Node process that can hold an OpenTelemetry span open for the duration of an LLM call. Convex actions are short, stateless, request/response functions — there's no "open a span, await a stream, close it later." This component is built for that shape: you call it once, with the finished input/output/timing/usage already in hand, and it does the rest.

Under the hood it exports a single completed OpenTelemetry span per call, using Langfuse's documented span attribute conventions (langfuse.observation.*, langfuse.trace.*) — no OpenTelemetry SDK dependency, just fetch. This matters more than it sounds: Langfuse's older /api/public/ingestion batch endpoint is deprecated and sunsets on Langfuse Cloud on November 16, 2026. Building against the OTLP endpoint from day one means this component doesn't need a rewrite in a few months.

Alongside the export, every trace, generation, and score is mirrored into Convex tables so you can query them reactively — build a live "recent generations" or "token usage" view without a separate call to Langfuse's API.

• One-shot logging — logGeneration() / logSpan() take a finished call and export it as a proper Langfuse trace + observation
• Scoring — recordScore() attaches an evaluation or user-rating score to any trace
• Reactive local queries — traces, observations, and scores are all queryable live from Convex
• No SDK dependency — talks to Langfuse's OTLP/HTTP and Scores REST endpoints directly over fetch

Table of Contents

• Install
• Quick Start
• Usage
• Continuing a Trace
• API Reference
• Type Reference
• Database Schema
• Testing
• Example App
• Limitations
• Troubleshooting
• Contributing
• Changelog

Install

npm install convex-langfuse

Requirements: Convex v1.33.1 or later, Node.js 18+, a Langfuse project (Cloud or self-hosted, OTel-compatible)

Quick Start

Three steps to send traces from your Convex app to Langfuse.

1. Add the component

In convex/convex.config.ts:

import { defineApp } from "convex/server";
import convexLangfuse from "convex-langfuse/convex.config";

const app = defineApp();
app.use(convexLangfuse);

export default app;

2. Set environment variables

npx convex env set LANGFUSE_PUBLIC_KEY pk-lf-xxxxxxxxxxxx
npx convex env set LANGFUSE_SECRET_KEY sk-lf-xxxxxxxxxxxx

3. Initialize the client

In convex/observability.ts:

import { components } from "./_generated/api";
import { Langfuse } from "convex-langfuse";

export const langfuse = new Langfuse(components.convexLangfuse, {
  publicKey: process.env.LANGFUSE_PUBLIC_KEY!,
  secretKey: process.env.LANGFUSE_SECRET_KEY!,
  // baseUrl defaults to the EU Langfuse Cloud region — set it for US/JP/self-hosted
});

Import langfuse from this file in any Convex action that calls an LLM.

Usage

Log a generation

export const chat = action({
  args: { userId: v.string(), prompt: v.string() },
  handler: async (ctx, args) => {
    const completion = await callYourModel(args.prompt);

    const { traceId } = await langfuse.logGeneration(ctx, {
      name: "chat-completion",
      userId: args.userId,
      model: "gpt-5",
      input: args.prompt,
      output: completion,
      inputTokens: completion.usage.inputTokens,
      outputTokens: completion.usage.outputTokens,
    });

    return { completion, traceId };
  },
});
// Returns: { traceId, observationId }

Log a non-LLM step

await langfuse.logSpan(ctx, {
  traceId,
  name: "vector-search",
  input: { query },
  output: { matches },
});

Record a score

export const rate = action({
  args: { traceId: v.string(), value: v.number() },
  handler: async (ctx, args) => {
    await langfuse.recordScore(ctx, {
      traceId: args.traceId,
      name: "user-rating",
      value: args.value, // e.g. 1 for thumbs up, 0 for thumbs down
    });
    return null;
  },
});

Read a trace reactively

export const getTrace = query({
  args: { traceId: v.string() },
  handler: async (ctx, args) => langfuse.getTrace(ctx, args),
});

export const getObservations = query({
  args: { traceId: v.string() },
  handler: async (ctx, args) => langfuse.listObservations(ctx, args),
});

Continuing a Trace

Pass the traceId returned from the first call into subsequent logGeneration()/logSpan() calls to group multiple observations (a retrieval step, then a generation, then a tool call) under one trace in the Langfuse UI:

const { traceId } = await langfuse.logSpan(ctx, { name: "retrieve-context", ... });
await langfuse.logGeneration(ctx, { traceId, name: "generate-answer", ... });

API Reference

| Method | Kind | Description | | --- | --- | --- | | logGeneration(ctx, args) | action | Exports a completed LLM call as a Langfuse generation | | logSpan(ctx, args) | action | Exports a completed non-LLM step as a Langfuse span | | recordScore(ctx, args) | action | Attaches an evaluation/rating score to a trace | | getTrace(ctx, args) | query | Fetch one trace by id | | listTracesByUser(ctx, args) | query | List a user's traces, newest first | | listTracesBySession(ctx, args) | query | List a session's traces, newest first | | listObservations(ctx, args) | query | List a trace's observations, oldest first | | listScores(ctx, args) | query | List a trace's scores | | listRecentTraces(ctx, args?) | query | Every trace, newest first, regardless of user/session — for a history view or dashboard | | getStats(ctx) | query | Aggregate trace/observation/score counts for a small dashboard |

Type Reference

type LogObservationArgs = {
  traceId?: string;       // omit to start a new trace
  traceName?: string;
  userId?: string;
  sessionId?: string;
  tags?: string[];
  release?: string;
  traceMetadata?: Record<string, unknown>;

  name: string;
  input?: unknown;
  output?: unknown;
  model?: string;
  inputTokens?: number;
  outputTokens?: number;
  level?: "DEBUG" | "DEFAULT" | "WARNING" | "ERROR";
  statusMessage?: string;
  metadata?: Record<string, unknown>;
  startTime?: number;      // ms epoch, defaults to now
  endTime?: number;        // ms epoch, defaults to startTime
};

type Trace = {
  traceId: string;
  name: string;
  userId?: string;
  sessionId?: string;
  tags?: string[];
  metadata?: string;
  release?: string;
};

type Observation = {
  observationId: string;
  traceId: string;
  type: "span" | "generation" | "event" | "embedding" | "agent" | "tool"
      | "chain" | "retriever" | "guardrail" | "evaluator";
  name: string;
  input?: string;
  output?: string;
  model?: string;
  inputTokens?: number;
  outputTokens?: number;
  level: "DEBUG" | "DEFAULT" | "WARNING" | "ERROR";
  statusMessage?: string;
  metadata?: string;
  startedAt: number;
  endedAt: number;
};

type Score = {
  scoreId: string;
  traceId: string;
  name: string;
  value: number;
  comment?: string;
};

Database Schema

traces: {
  traceId, name, userId?, sessionId?, tags?, metadata?, release?,
  createdAt, updatedAt,
}

observations: {
  observationId, traceId, type, name, input?, output?, model?,
  inputTokens?, outputTokens?, level, statusMessage?, metadata?,
  startedAt, endedAt, createdAt,
}

scores: {
  scoreId, traceId, name, value, comment?, createdAt,
}

listRecentTraces() and getStats() read across these tables directly with a full scan — fine for a history view or small dashboard, not intended as a high-volume production query path.

Testing

npm run test

Component logic is tested with convex-test in src/component/lib.test.ts. Import convex-langfuse/test in your own app to register this component's schema against your test instance.

Example App

example/ is a full Vite + React demo that exercises the entire component end to end against your own Langfuse account:

• Log a generation — simulate a finished chat completion (model, prompt, completion, token counts, tags) and export it as a Langfuse generation.
• Continue the trace — attach a non-LLM step (a retrieval, a tool call, a guardrail check) to the same trace via logSpan().
• Rate the response — attach a 👍/👎 recordScore() call, with an optional comment.
• Trace explorer — look up any trace by id, or browse the traces recorded for a given user/session, all reactive.
• History — every trace ever recorded in the deployment, newest first, independent of the browser-local demo identity that resets on reload. Expand a trace to see its full generation/step/score detail, and hit Replay this to recreate the whole scenario as a brand-new trace live, without retyping anything — useful for anyone poking at the demo who wants to see the flow without filling out forms themselves.
• Activity console — a side-docked live log of every call this app makes into the Langfuse client. Since Langfuse ingestion is one-way (no webhooks call back into Convex), this is the closest thing to a live feed the example has.

Run it from the repo root:

npm install --legacy-peer-deps
npx convex env set LANGFUSE_PUBLIC_KEY pk-lf-xxxxxxxxxxxx
npx convex env set LANGFUSE_SECRET_KEY sk-lf-xxxxxxxxxxxx
npm run dev

npm run dev starts the Convex backend and the Vite frontend together — there's no need to cd example or run either one separately.

Limitations

• No live/streaming spans — each call exports one already-finished observation. For a call you want to see "in progress," log it once it completes rather than trying to open and later close a span.
• Trace-level fields (userId, sessionId, tags, metadata) are attached to every span you log under that trace, per Langfuse's own OTel guidance for filterable trace attributes — there's no separate "create trace" call.
• This component does not vendor the Langfuse SDK or an OpenTelemetry SDK; it constructs minimal OTLP/HTTP JSON payloads directly, which keeps it dependency-free but means it only covers the attributes this README documents.

Troubleshooting

Nothing shows up in the Langfuse UI — confirm baseUrl matches your project's region (EU, US, JP, HIPAA, or your self-hosted URL) and that the OTLP endpoint isn't blocked by a firewall between Convex and Langfuse.

401 from Langfuse — publicKey/secretKey are combined into a Basic auth header; confirm both are set and belong to the same project.

Generation shows no cost/usage — cost is computed by Langfuse from inputTokens/outputTokens and the model's pricing table; confirm model matches a model name Langfuse recognizes for your provider.

Contributing

See CONTRIBUTING.md.

Changelog

See CHANGELOG.md.