# Langfuse Prompt Management — Versioned Prompts + A/B Tests

> Langfuse Prompt Management versions, labels, and A/B tests prompts. Edit in UI, fetch via SDK, swap models without code deploys.

## Install

Copy the content below into your project:

## Quick Use

1. `pip install langfuse`
2. Set LANGFUSE_PUBLIC_KEY / SECRET_KEY / HOST in env
3. Author prompt in UI, label `production`, fetch with `lf.get_prompt(name, label='production')`

---

## Intro

Langfuse Prompt Management stores prompts as versioned, labeled artifacts in Langfuse — edit in the UI, fetch from the SDK at runtime, swap models or templates without code deploys. Each prompt has a label (production, staging, experiment-v3) you point your code at. Best for: teams iterating on prompts faster than they can ship code, A/B testing prompts, granting non-engineers safe edit access. Works with: Python and JS SDKs, LangChain Hub-compatible, OpenAI and Anthropic message format. Setup time: 5 minutes.

---

### Push a prompt programmatically

```python
from langfuse import Langfuse
lf = Langfuse()  # picks up LANGFUSE_PUBLIC_KEY / SECRET_KEY / HOST

lf.create_prompt(
    name="support-triage",
    prompt=[
        {"role": "system", "content": "You triage support tickets into urgent / billing / general."},
        {"role": "user",   "content": "Ticket: {{ticket_text}}"},
    ],
    config={"model": "claude-3-5-sonnet-20241022", "temperature": 0.2},
    labels=["production"],
)
```

### Fetch + execute (with caching)

```python
from langfuse import Langfuse
from langfuse.openai import openai  # auto-tracing wrapper

lf = Langfuse()
prompt = lf.get_prompt("support-triage", label="production", cache_ttl_seconds=60)

messages = prompt.compile(ticket_text="My card was charged twice for order #4521")
resp = openai.chat.completions.create(
    model=prompt.config["model"],
    messages=messages,
    temperature=prompt.config["temperature"],
)
```

### A/B testing two prompt versions

Tag two versions with separate labels (`production-a`, `production-b`), split traffic in code by user_id hash, then compare success metrics in the Langfuse Scores tab.

```python
import hashlib

def variant(user_id: str) -> str:
    return "production-a" if int(hashlib.md5(user_id.encode()).hexdigest(), 16) % 2 == 0 else "production-b"

prompt = lf.get_prompt("support-triage", label=variant(user_id))
```

### Why labels not version pins

| Approach | Behavior |
|---|---|
| `label="production"` | Latest production-tagged version. Edits in UI go live without code deploy. |
| `version=7` | Hard pin. Code change required to move forward. Use only for compliance freezes. |

### Self-hosted Langfuse

```bash
git clone https://github.com/langfuse/langfuse
cd langfuse
docker compose up -d
# Open http://localhost:3000, create org + project, copy keys into .env
```

---

### FAQ

**Q: How is this different from LangChain Hub?**
A: Langfuse Prompt Management ships with full traces — you see which prompt version produced which output, with cost and latency. LangChain Hub is registry-only with no observability. Langfuse also self-hosts; Hub is LangSmith-tied.

**Q: Can non-engineers edit prompts safely?**
A: Yes — give them Langfuse project access with the Editor role. They edit in UI, save creates a new version. Production label only moves when an admin promotes it, so junior edits can't ship without review.

**Q: What's the cache_ttl_seconds default?**
A: 60 seconds. The SDK fetches and caches; subsequent calls within 60s use the cache. Tune lower for fast-iteration dev (1s) or higher for stable prod (300s+) to reduce control-plane load.

---

## Source & Thanks

> Built by [Langfuse](https://github.com/langfuse). Licensed under MIT.
>
> [langfuse/langfuse](https://github.com/langfuse/langfuse) — ⭐ 8,000+

---

<!-- ZH -->

## 快速使用

1. `pip install langfuse`
2. 设 LANGFUSE_PUBLIC_KEY / SECRET_KEY / HOST 环境变量
3. UI 编辑 prompt 打 `production` 标签，用 `lf.get_prompt(name, label='production')` 拉取

---

## 简介

Langfuse Prompt Management 把 prompt 存成 Langfuse 里的版本化、带标签的工件 —— UI 编辑、运行时从 SDK 拉、换模型或模板不用发版。每个 prompt 有标签（production / staging / experiment-v3），代码指向哪个。适合 prompt 迭代速度比代码发版快的团队、A/B 测 prompt、给非工程师安全编辑权限。兼容 Python / JS SDK、LangChain Hub 格式、OpenAI 和 Anthropic message 格式。装机时间 5 分钟。

---

### 程序化推 prompt

```python
from langfuse import Langfuse
lf = Langfuse()  # 读 LANGFUSE_PUBLIC_KEY / SECRET_KEY / HOST

lf.create_prompt(
    name="support-triage",
    prompt=[
        {"role": "system", "content": "把客服工单分流成 urgent / billing / general 三类。"},
        {"role": "user",   "content": "工单：{{ticket_text}}"},
    ],
    config={"model": "claude-3-5-sonnet-20241022", "temperature": 0.2},
    labels=["production"],
)
```

### 拉取 + 执行（带缓存）

```python
from langfuse import Langfuse
from langfuse.openai import openai  # 自动追踪 wrapper

lf = Langfuse()
prompt = lf.get_prompt("support-triage", label="production", cache_ttl_seconds=60)

messages = prompt.compile(ticket_text="订单 #4521 我的卡被刷了两次")
resp = openai.chat.completions.create(
    model=prompt.config["model"],
    messages=messages,
    temperature=prompt.config["temperature"],
)
```

### 两版 prompt A/B 测

给两版打不同标签（`production-a` / `production-b`），代码里按 user_id hash 分流，在 Langfuse Scores 标签页对比成功指标。

```python
import hashlib

def variant(user_id: str) -> str:
    return "production-a" if int(hashlib.md5(user_id.encode()).hexdigest(), 16) % 2 == 0 else "production-b"

prompt = lf.get_prompt("support-triage", label=variant(user_id))
```

### 为啥用标签不用版本号锁定

| 方式 | 行为 |
|---|---|
| `label="production"` | 拿当前 production 标签的最新版。UI 编辑直接生效，不发版。|
| `version=7` | 硬锁。要推进必须改代码。仅合规冻结期用。|

### 自托管 Langfuse

```bash
git clone https://github.com/langfuse/langfuse
cd langfuse
docker compose up -d
# 打开 http://localhost:3000，建组织+项目，把 key 复制到 .env
```

---

### FAQ

**Q: 跟 LangChain Hub 啥区别？**
A: Langfuse Prompt Management 自带完整 trace —— 哪个 prompt 版本产出哪个结果，附成本和延迟。LangChain Hub 仅注册表，没观测。Langfuse 还能自托管，Hub 绑 LangSmith。

**Q: 非工程师能安全编辑 prompt 吗？**
A: 能 —— 给他们 Langfuse 项目 Editor 角色。UI 编辑，保存即新建版本。production 标签只能由 admin 推进，所以初级编辑不评审上不了线。

**Q: cache_ttl_seconds 默认多少？**
A: 60 秒。SDK 拉取并缓存；60 秒内的后续调用用缓存。开发态调小（1 秒）快迭代，生产态调大（300+）减控制面压力。

---

## 来源与感谢

> Built by [Langfuse](https://github.com/langfuse). Licensed under MIT.
>
> [langfuse/langfuse](https://github.com/langfuse/langfuse) — ⭐ 8,000+


---
Source: https://tokrepo.com/en/workflows/langfuse-prompt-management-versioned-prompts-a-b-tests
Author: Langfuse