Tune the gateway for AI traffic¶
Large language model (LLM) and Model Context Protocol (MCP) traffic differs from representational state transfer (REST) traffic in ways the chart defaults don't anticipate. Requests stay open for tens of seconds, bodies run to megabytes, and guardrails inspect every one of those bodies. This page covers the settings that matter because of those differences.
Raise the timeouts for long completions¶
The chart allows 60 seconds for a complete request-to-response cycle on a route. A large completion, a reasoning model, or a long tool-use chain regularly exceeds that, and the client receives a gateway timeout partway through a response the provider was still producing.
gateway:
config:
router:
upstream:
timeouts:
route_timeout_ms: 300000 # 5 minutes for a full completion
route_idle_timeout_ms: 300000 # Gap allowed between stream chunks
connect_timeout_ms: 5000
route_idle_timeout_ms is what keeps a streaming response alive: it bounds the gap between chunks arriving from the provider, not the length of the response as a whole. Set it longer than the longest pause you expect between tokens.
Note
Raise the matching timeouts on whatever sits in front of the gateway as well. An ingress controller or load balancer with a 60-second read timeout cuts the response off regardless of what the gateway allows.
Settings that the structured chart values don't expose can be appended as raw TOML through gateway.config_toml, which the chart merges into the generated config.toml. The token cost tracking section below uses this mechanism.
Give guardrails room to run¶
Guardrails execute in the policy engine on the request path, the response path, or both. Two timeouts bound them, and both apply per request.
gateway:
config:
router:
policy_engine:
timeout_ms: 60000
message_timeout_ms: 60000
route_cache_action: RETAIN
Guardrails that call an external service—Azure Content Safety, AWS Bedrock guardrails, or a semantic prompt guardrail that generates embeddings—spend most of that budget on the network call. Keep these values below route_timeout_ms so the route timeout stays the outer bound.
Leave route_cache_action at RETAIN, which keeps the route cache warm across requests.
Guardrails that reach an external service add that service's latency and its failure modes to every request. Deploy the guardrail service in the same region as the gateway, and check what your chosen guardrail does when the service is unreachable before you rely on it in production.
Supply pricing data for token cost tracking¶
The llm_cost_v1 policy converts token counts into cost using a pricing file. The chart mounts a bundled model_prices.json at /etc/policy-engine/llm-pricing/model_prices.json:
gateway:
gatewayRuntime:
policies:
llmPricing:
enabled: true
configMapName: "" # Empty uses the bundled pricing data
The bundled file is fixed at the release you deploy, so it doesn't reflect later provider price changes, and it has no entry for a self-hosted or private model. Supply your own by creating a ConfigMap with a model_prices.json key and naming it in configMapName:
kubectl create configmap llm-model-prices \
--namespace <your-namespace> \
--from-file=model_prices.json=./model_prices.json
When you name a ConfigMap, the chart doesn't create the built-in one. Point the policy at the mounted path through the raw configuration passthrough:
gateway:
config_toml: |
[policy_configurations.llm_cost_v1]
pricing_file = "/etc/policy-engine/llm-pricing/model_prices.json"
Cost figures are only as accurate as this file. Treat it as data to review on a schedule rather than as something to set once.
Provision infrastructure for semantic caching¶
Semantic caching depends on two external services that the gateway chart doesn't deploy, so provision both before you enable the policy:
- A vector database — Redis or Milvus — that stores the cached responses and their embeddings. Size it for your retention window, place it in the same region as the gateway, and secure it: it holds prompt and completion content in full.
- An embedding provider — OpenAI, Mistral, or Azure OpenAI — called on every request that reaches the policy. Its latency is added to every cache miss, and its API key is stored as a gateway secret, encrypted with the key from Security hardening.
The cache is shared state. Every gateway runtime replica pointed at the same vector database serves from and writes to one cache, which is what you want across replicas of a single environment. Use separate databases or indexes for separate environments, so a staging prompt can't be served to a production caller.
Log and trace the AI traffic¶
The chart ships the controller log level at debug. Move it to info in production, and keep JSON formatting so a log aggregator can parse it:
gateway:
config:
controller:
logging:
level: info
format: json
policy_engine:
logging:
level: info
format: json
Prompts and completions are sensitive
Debug-level logging can record request and response bodies, which on an AI Gateway means user prompts and model completions. Before you enable debug-level logging in production, confirm that your log retention and access controls suit that content. If you forward traffic data to an external analytics service, control what leaves the gateway with the analytics header filter. For where the logs go and how to read them, see Gateway logging.