fika-langwatch 0.2.3

LangChain wrapper with automatic fallback and alert notifications when API keys fail

fika-langwatch

LangChain wrapper with automatic fallback and alert notifications when API keys fail.

Copyright (c) 2026 FIKA Private Limited. All Rights Reserved.

Features

Feature	Description
Automatic Fallback	When a model fails, automatically try the next one in the chain
Smart Key Skipping	Skip unhealthy keys to go straight to fallback (configurable timeout)
Alert Notifications	Get notified via Email, Slack, or Webhooks when keys fail
Rate Limiting	Built-in cooldown to prevent alert spam (in-memory, no Redis needed)
Tool Binding	Call .bind_tools() once, applies to ALL underlying models
Sync & Async	Supports both .invoke() and .ainvoke()
Health Tracking	Track key health status with automatic recovery detection
Per-Key Alerts	Alert on EACH key failure with full details (provider, model, masked key, app name)

Installation

pip install fika-langwatch

# With optional dependencies
pip install fika-langwatch[email]      # Email alerts (aiosmtplib)
pip install fika-langwatch[slack]      # Slack alerts (httpx)
pip install fika-langwatch[webhook]    # Webhook alerts (httpx)
pip install fika-langwatch[all]        # All alert channels

# Provider-specific
pip install fika-langwatch[google]     # Google Gemini
pip install fika-langwatch[openai]     # OpenAI
pip install fika-langwatch[anthropic]  # Anthropic Claude
pip install fika-langwatch[providers]  # All providers

Quick Start

Option 1: Config-based (Recommended)

from langwatch import ChatWithFallback
from langwatch.alerts import EmailAlert, SlackAlert
from langchain_core.messages import HumanMessage

# Create with config - models are created automatically
chat = ChatWithFallback.from_config(
    models=[
        {
            "name": "gemini-1",
            "provider": "google",
            "model": "gemini-2.5-flash",
            "api_key": "AIza...",
        },
        {
            "name": "gemini-2",
            "provider": "google",
            "model": "gemini-2.5-flash",
            "api_key": "AIza...",
        },
        {
            "name": "fallback",
            "provider": "openrouter",
            "model": "x-ai/grok-4.1-fast",
            "api_key": "sk-...",
            "is_fallback": True,
        },
    ],
    alerts=[
        EmailAlert(
            smtp_server="smtp.gmail.com",
            smtp_port=587,
            username="[e-mail]",
            password="app-password",
            to=["[e-mail]"],
        ),
        SlackAlert(webhook_url="https://hooks.slack.com/services/..."),
    ],
    app_name="MyApp",         # Shows in alerts: [MyApp] API Key Failure
    cooldown_seconds=300,     # Per-key cooldown (default: 5 minutes)
    skip_unhealthy=True,      # Skip failed keys (default: True)
    unhealthy_timeout=300,    # Retry failed keys after 5 minutes (default: 300s)
)

# Bind tools - applies to ALL models
chat_with_tools = chat.bind_tools([your_tool_1, your_tool_2])

# Use like any LangChain model (sync or async)
response = chat_with_tools.invoke([HumanMessage(content="Hello!")])
response = await chat_with_tools.ainvoke([HumanMessage(content="Hello!")])

Option 2: Manual Models (Full Flexibility)

from langwatch import ChatWithFallback
from langwatch.alerts import EmailAlert
from langchain_google_genai import ChatGoogleGenerativeAI
from langchain_openai import ChatOpenAI

# Create your own models - works with ANY LangChain-compatible model
models = [
    ChatGoogleGenerativeAI(model="gemini-2.5-flash", google_api_key="..."),
    ChatGoogleGenerativeAI(model="gemini-2.5-flash", google_api_key="..."),
    ChatOpenAI(model="grok-4.1", base_url="https://openrouter.ai/api/v1", api_key="..."),
]

chat = ChatWithFallback(
    models=models,
    model_names=["gemini-1", "gemini-2", "fallback"],
    alerts=[EmailAlert(...)],
)

# Bind tools and use
chat_with_tools = chat.bind_tools(tools)
response = await chat_with_tools.ainvoke(messages)

---

Feature Details

1. Automatic Fallback

When a model fails (rate limit, API error, etc.), the system automatically tries the next model in the chain.

# Define multiple models - they're tried in order
models=[
    {"name": "primary-1", ...},
    {"name": "primary-2", ...},
    {"name": "fallback", ..., "is_fallback": True},
]

2. Smart Key Skipping (skip_unhealthy)

By default, unhealthy keys are skipped to avoid wasting time on failed APIs. After a timeout, they're retried automatically.

chat = ChatWithFallback.from_config(
    models=[...],
    skip_unhealthy=True,      # Skip unhealthy keys (default: True)
    unhealthy_timeout=300,    # Retry after 5 minutes (default: 300 seconds)
)

Behavior with skip_unhealthy=True:

Request 1: Gemini-1 ❌ → Gemini-2 ❌ → Gemini-3 ❌ → Fallback ✅ (alert sent)
Request 2: Skip Gemini-1,2,3 → Fallback ✅ (fast! no wasted API calls)
Request 3: Skip Gemini-1,2,3 → Fallback ✅
...
After 5 min timeout:
Request N: Gemini-1 (retry) ✅ → Done! (key recovered)

Behavior with skip_unhealthy=False:

Request 1: Gemini-1 ❌ → Gemini-2 ❌ → Gemini-3 ❌ → Fallback ✅
Request 2: Gemini-1 ❌ → Gemini-2 ❌ → Gemini-3 ❌ → Fallback ✅ (slow, retries all)

3. Alert Notifications

Get notified when all primary keys fail and fallback is activated.

Email Alerts

from langwatch.alerts import EmailAlert

alert = EmailAlert(
    smtp_server="smtp.gmail.com",
    smtp_port=587,
    username="[e-mail]",
    password="your-app-password",  # Use Gmail App Password
    to=["[e-mail]"],
    cc=["[e-mail]"],       # Optional
    bcc=["[e-mail]"],      # Optional
    from_name="LangWatch Alerts",  # Optional (default: "LangWatch Alerts")
    use_tls=True,                  # Optional (default: True)
)

Slack Alerts

from langwatch.alerts import SlackAlert

alert = SlackAlert(
    webhook_url="https://hooks.slack.com/services/T.../B.../xxx",
    channel="#alerts",              # Optional - override webhook default
    username="LangWatch Bot",       # Optional
    icon_emoji=":warning:",         # Optional
)

Webhook Alerts (Generic HTTP)

from langwatch.alerts import WebhookAlert

alert = WebhookAlert(
    url="https://your-api.com/alerts",
    headers={"Authorization": "Bearer token"},
    method="POST",                  # Optional (default: "POST")
    timeout=10.0,                   # Optional (default: 10 seconds)
)

4. Rate Limiting (Alert Cooldown)

Prevents alert spam by limiting how often alerts are sent.

chat = ChatWithFallback.from_config(
    models=[...],
    alerts=[...],
    cooldown_seconds=3600,  # Only 1 alert per hour (default: 3600)
)

5. Tool Binding

Bind tools once - applies to ALL underlying models automatically.

from langchain_core.tools import tool

@tool
def search_knowledge_base(query: str) -> str:
    """Search the knowledge base."""
    return "Results..."

@tool
def book_appointment(date: str, time: str) -> str:
    """Book an appointment."""
    return "Booked!"

# Bind tools to ALL models at once
chat_with_tools = chat.bind_tools([search_knowledge_base, book_appointment])

# Now all models (primary and fallback) have these tools bound
response = await chat_with_tools.ainvoke(messages)

6. Callbacks

Get notified programmatically when keys fail or fallback activates.

def on_key_failure(key_name: str, error: str):
    """Called when any key fails."""
    print(f"Key {key_name} failed: {error}")
    # Log to your monitoring system, etc.

def on_fallback_activated(fallback_key: str):
    """Called when fallback is activated."""
    print(f"Now using fallback: {fallback_key}")
    # Trigger additional alerts, update dashboards, etc.

chat = ChatWithFallback.from_config(
    models=[...],
    alerts=[...],
    on_key_failure=on_key_failure,
    on_fallback_activated=on_fallback_activated,
)

7. Health Status Monitoring

Check the current status of all keys programmatically.

status = chat.get_status()
print(status)
# {
#     "total_keys": 4,
#     "healthy_keys": 1,
#     "failed_keys": 3,
#     "all_primary_failed": True,
#     "keys": [
#         {"name": "gemini-1", "is_healthy": False, "failure_count": 5, ...},
#         {"name": "gemini-2", "is_healthy": False, "failure_count": 3, ...},
#         {"name": "gemini-3", "is_healthy": False, "failure_count": 2, ...},
#         {"name": "fallback", "is_healthy": True, "failure_count": 0, ...},
#     ]
# }

8. Per-Key Alerts with App Name

Alerts are sent on EVERY key failure with full details. Use app_name to identify which client/app the alert is from.

chat = ChatWithFallback.from_config(
    models=[...],
    alerts=[...],
    app_name="ClientA",       # Shows in alert subject: [ClientA] API Key Failure
    cooldown_seconds=300,     # Per-key cooldown (default: 5 minutes)
)

Alert example:

[ClientA] API Key Failure - gemini-1

[ClientA] API key 'gemini-1' (primary) has failed.
Provider: google
Model: gemini-2.5-flash
API Key: AIza...Xyz9
Failure Count: 3
Error: Rate limit exceeded. Please try again in 60 seconds.

Per-key cooldown prevents spam:

gemini-1 fails → alert sent
gemini-1 fails again → NO alert (in 5 min cooldown)
gemini-2 fails → alert sent (different key)
After 5 min: gemini-1 fails → alert sent

---

Supported Providers

When using from_config(), these providers are auto-created:

Provider	Value	LangChain Class	Notes
Google Gemini	"google"	ChatGoogleGenerativeAI	Requires langchain-google-genai
OpenAI	"openai"	ChatOpenAI	Requires langchain-openai
Anthropic Claude	"anthropic"	ChatAnthropic	Requires langchain-anthropic
OpenRouter	"openrouter"	ChatOpenAI	Uses OpenRouter base_url

For other providers (Grok, Mistral, etc.), create the model manually and pass it directly.

---

Configuration Reference

ChatWithFallback.from_config() Parameters

Parameter	Type	Default	Description
models	List[dict]	Required	List of model configurations
alerts	List[AlertChannel]	[]	Alert channels (Email, Slack, Webhook)
app_name	str	None	App/client name for alert subject (e.g., "[MyApp] API Key Failure")
cooldown_seconds	int	300	Per-key cooldown between alerts (5 minutes default)
skip_unhealthy	bool	True	Skip unhealthy keys until timeout
unhealthy_timeout	int	300	Seconds before retrying unhealthy keys
on_key_failure	Callable	None	Callback when a key fails
on_fallback_activated	Callable	None	Callback when fallback activates

Model Configuration

{
    "name": "gemini-1",           # Unique name for alerts/logging
    "provider": "google",         # Provider: google, openai, anthropic, openrouter
    "model": "gemini-2.5-flash",  # Model name
    "api_key": "AIza...",         # API key
    "is_fallback": False,         # True for fallback models (default: False)
    "extra_config": {             # Optional provider-specific config
        "temperature": 0.7,
        "max_retries": 0,
    }
}

---

How It Works

┌─────────────────────────────────────────────────────────────────┐
│                         Request                                  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│  Is key healthy OR timeout expired?                              │
│  ┌─────────┐    ┌─────────┐    ┌─────────┐    ┌─────────┐      │
│  │Gemini-1 │───▶│Gemini-2 │───▶│Gemini-3 │───▶│Fallback │      │
│  │(skip?)  │    │(skip?)  │    │(skip?)  │    │         │      │
│  └─────────┘    └─────────┘    └─────────┘    └─────────┘      │
└─────────────────────────────────────────────────────────────────┘
                              │
              ┌───────────────┴───────────────┐
              │                               │
         Success ✅                       Failure ❌
              │                               │
              ▼                               ▼
┌─────────────────────┐         ┌─────────────────────────────────┐
│  Mark key healthy   │         │  Mark key unhealthy             │
│  Return response    │         │  Try next key                   │
└─────────────────────┘         │  If entering fallback → Alert   │
                                └─────────────────────────────────┘

---

License

Copyright (c) 2026 FIKA Private Limited. All Rights Reserved.

This is proprietary software. Unauthorized copying, modification, or distribution is prohibited.

Authors:

• Rahul Kumar - [e-mail]
• Nikhil Sukhte - [e-mail]

For licensing inquiries, contact: [e-mail]