The TensorZero Gateway offers multiple strategies to handle errors and improve reliability. These strategies are defined at three levels: models (model provider routing), variants (variant retries), and functions (variant fallbacks). You can combine these strategies to define complex fallback behavior.

​
Model Provider Routing

We can specify that a model is available on multiple providers using its routing field. If we include multiple providers on the list, the gateway will try each one sequentially until one succeeds or all fail. In the example below, the gateway will first try OpenAI, and if that fails, it will try Azure. 
[models.gpt_4o_mini]
# Try the following providers in order:
# 1. `models.gpt_4o_mini.providers.openai`
# 2. `models.gpt_4o_mini.providers.azure`
routing = ["openai", "azure"]

[models.gpt_4o_mini.providers.openai]
type = "openai"
model_name = "gpt-4o-mini-2024-07-18"

[models.gpt_4o_mini.providers.azure]
type = "azure"
deployment_id = "gpt4o-mini-20240718"
endpoint = "https://your-azure-openai-endpoint.openai.azure.com"

[functions.extract_data]
type = "chat"

[functions.extract_data.variants.gpt_4o_mini]
type = "chat_completion"
model = "gpt_4o_mini"

​
Variant Retries

We can add a retries field to a variant to specify the number of times to retry that variant if it fails. The retry strategy is a truncated exponential backoff with jitter. In the example below, the gateway will retry the variant four times (i.e. a total of five attempts), with a maximum delay of 10 seconds between retries. 
[functions.extract_data]
type = "chat"

[functions.extract_data.variants.claude_3_5_haiku]
type = "chat_completion"
model = "anthropic::claude-3-5-haiku-20241022"
# Retry the variant up to four times, with a maximum delay of 10 seconds between retries.
retries = { num_retries = 4, max_delay_s = 10 }

​
Variant Fallbacks

If we specify multiple variants for a function, the gateway will try different variants until one succeeds or all fail. The sampling behavior depends on how the weights are specified:
  • If no weights are specified for any variants, the gateway will sample between them uniformly.
  • If a variant’s weight is set to zero, it will never be sampled unless explicitly pinned at inference time using variant_name.
  • If you mix variants with positive and unspecified weights, the gateway will sample the positive weighted variants first, and only use the unspecified weighted variants as fallbacks.
In the example below, the gateway will first sample and attempt the variants with positive weights (gpt_4o_mini or claude_3_5_haiku). If all of those variants fail, the gateway will sample and attempt the variants with unspecified weights (gemini_1_5_flash_8b or ministral_8b). The gateway will never sample the variants with zero weights (ministral_8b), unless explicitly pinned at inference time. 
[functions.extract_data]
type = "chat"

[functions.extract_data.variants.gpt_4o_mini]
type = "chat_completion"
model = "openai::gpt-4o-mini-2024-07-18"
weight = 0.7

[functions.extract_data.variants.claude_3_5_haiku]
type = "chat_completion"
model = "anthropic::claude-3-5-haiku-20241022"
weight = 0.3

[functions.extract_data.variants.gemini_1_5_flash_8b]
type = "chat_completion"
model = "google_ai_studio_gemini::gemini-1.5-flash-8b"

[functions.extract_data.variants.grok_2]
type = "chat_completion"
model = "xai::grok-2-1212"

[functions.extract_data.variants.ministral_8b]
type = "chat_completion"
model = "mistral::ministral-8b-2410"
weight = 0

​
Combining Strategies

We can combine strategies to define complex fallback behavior. The gateway will try the following strategies in order:
  1. Model Provider Routing
  2. Variant Retries
  3. Variant Fallbacks
In other words, the gateway will follow a strategy like the pseudocode below. 
while variants:
    # First sample variants with non-zero weight, then variants with zero weight
    variant = sample_variant(variants)  # sampling without replacement

    for _ in range(num_retries + 1):
        for provider in variant.routing:
            try:
                return inference(variant, provider)
            except:
                continue

​
Load Balancing

TensorZero doesn’t currently offer an explicit strategy for load balancing API keys, but you can achieve a similar effect by defining multiple variants with appropriate weights. We plan to add a streamlined load balancing strategy in the future. In the example below, the gateway will split the traffic between two variants (gpt_4o_mini_api_key_A and gpt_4o_mini_api_key_B). Each variant leverages a model with providers that use different API keys (OPENAI_API_KEY_A and OPENAI_API_KEY_B). See Credential Management for more details on credential management. 
[models.gpt_4o_mini_api_key_A]
routing = ["openai"]

[models.gpt_4o_mini_api_key_A.providers.openai]
type = "openai"
model_name = "gpt-4o-mini-2024-07-18"
api_key_location = "env:OPENAI_API_KEY_A"

[models.gpt_4o_mini_api_key_B]
routing = ["openai"]

[models.gpt_4o_mini_api_key_B.providers.openai]
type = "openai"
model_name = "gpt-4o-mini-2024-07-18"
api_key_location = "env:OPENAI_API_KEY_B"

[functions.extract_data]
type = "chat"

[functions.extract_data.variants.gpt_4o_mini_api_key_A]
type = "chat_completion"
model = "openai::gpt-4o-mini-2024-07-18"
weight = 0.5

[functions.extract_data.variants.gpt_4o_mini_api_key_B]
type = "chat_completion"
model = "openai::gpt-4o-mini-2024-07-18"
weight = 0.5

​
Timeouts

You can set granular timeouts for individual requests to a model provider, model, or variant using the timeouts field in the corresponding configuration block. You can define timeouts for non-streaming and streaming requests separately: timeouts.non_streaming.total_ms corresponds to the total request duration and timeouts.streaming.ttft_ms corresponds to the time to first token (TTFT). For example, the following configuration sets a 15-second timeout for non-streaming requests and a 3-second timeout for streaming requests (TTFT) to a particular model provider. 
[models.model_name.providers.provider_name]
# ...
timeouts = { non_streaming.total_ms = 15000, streaming.ttft_ms = 3000 }
# ...
This setting applies to individual requests to the model provider. If you’re using an advanced variant type that performs multiple requests, the timeout will apply to each request separately. If you’ve defined retries and fallbacks, the timeout will apply to each retry and fallback separately. This setting is particularly useful if you’d like to retry or fallback on a request that’s taking too long. If you specify timeouts for a model, they apply to every inference request in the model’s scope, including retries and fallbacks. If you specify timeouts for a variant, they apply to every inference request in the variant’s scope, including retries and fallbacks. For advanced variant types that perform multiple requests, the timeout applies collectively to the sequence of all requests. Separately, you can set a global timeout for the entire inference request using the TensorZero client’s timeout field (or simply killing the request if you’re using a different client).

​
Technical Notes

  • For variant types that require multiple model inferences (e.g. best-of-N sampling), the routing fallback applies to each individual model inference separately.