Function Calling Across Providers: A Schema-Safe Developer Guide

Function Calling Across Providers: A Schema-Safe Developer Guide#

What is function calling across providers?#

Function calling lets a model propose structured arguments for an application-defined tool. Cross-provider support becomes difficult because names, schemas, parallel calls, streaming events, and refusal behavior differ. The stable boundary should be your own internal tool format, with provider adapters translating into it.

A portable tool contract#

Define a strict JSON Schema. Validate types, ranges, enums, and authorization server-side. Return tool results as data, not executable instructions. For destructive or financial operations, require an explicit user confirmation even when the model’s arguments are valid.

tools = [{"type":"function","function":{
  "name":"lookup_order",
  "description":"Read an order status",
  "parameters":{"type":"object","properties":{"order_id":{"type":"string"}},"required":["order_id"],"additionalProperties":False}
}}]
response = client.chat.completions.create(model="MODEL_ID", messages=messages, tools=tools)

Provider comparison and pricing#

Retry the tool execution only when it is idempotent. A timeout after a successful payment call must not trigger a second payment. Store a call ID and use an idempotency key.

Why this topic matters to developers#

AI integrations fail less often when the application treats a model as a replaceable service rather than a hard-coded vendor feature. The useful unit is a request contract: inputs, outputs, latency expectations, safety rules, and a cost ceiling. That contract makes it possible to test a model directly, route through a gateway, and change providers without rewriting the product.

The examples below use an OpenAI-compatible endpoint. Replace the model identifier with the exact model exposed in your account and check the provider's current documentation before deploying. Model names, limits, and prices change; a resilient integration should discover capabilities and record the provider response rather than assuming that a blog post is a billing contract.

Comparison: direct provider, hosted tool, or Crazyrouter#

Quick-start API pattern#

cURL#

export CR_API_KEY='replace-with-your-key'
curl https://crazyrouter.com/v1/chat/completions \
  -H "Authorization: Bearer $CR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"MODEL_ID","messages":[{"role":"user","content":"Return a concise JSON health check."}],"temperature":0.2}'

Python#

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["CR_API_KEY"],
    base_url="https://crazyrouter.com/v1",
)
response = client.chat.completions.create(
    model="MODEL_ID",
    messages=[{"role": "user", "content": "Explain the result in three bullets."}],
    timeout=45,
)
print(response.choices[0].message.content)

Node.js#

import OpenAI from "openai";
const client = new OpenAI({
  apiKey: process.env.CR_API_KEY,
  baseURL: "https://crazyrouter.com/v1"
});
const result = await client.chat.completions.create({
  model: "MODEL_ID",
  messages: [{ role: "user", content: "Return a short deployment checklist." }]
});
console.log(result.choices[0].message.content);

Use environment variables or a secret manager; never commit a key. Add request IDs, timeouts, bounded retries, and structured logs before moving this snippet into a queue worker.

Production rollout checklist#

Start with a shadow test against recorded, consented examples. Define an acceptance rubric before looking at outputs: correctness, format compliance, latency, safety, and cost. Then release to a small percentage of traffic with a kill switch. Keep the previous route available until the new one has survived peak load and a provider incident.

For observability, record a correlation ID, tenant, model and route, sanitized prompt hash, token or media usage, queue time, inference time, finish reason, error class, and estimated cost. Do not log raw confidential prompts by default. Build dashboards for p50/p95 latency, timeout rate, schema-validation failures, retry amplification, and spend per accepted result. These measurements make provider comparisons reproducible and reveal regressions that a manual demo will miss.

Frequently asked questions#

Is function calling across providers available through an API?#

Availability depends on the current model catalog, account, region, and route. Check the live documentation and send a small test request before committing to an architecture.

Is Crazyrouter cheaper than the official provider?#

Not automatically. Compare the current rate card and your effective cost, including retries, engineering work, storage, and accepted-output rate. Crazyrouter is useful when portability, centralized routing, and one compatible endpoint matter.

How should I handle failures?#

Set a timeout, classify 4xx versus 5xx errors, retry only transient failures with exponential backoff, and use an idempotency key for asynchronous or side-effecting operations.

Summary#

The practical way to adopt function calling across providers is to start with a narrow benchmark, normalize the request contract, and measure quality, latency, and cost together. For a faster multi-model starting point, review the current Crazyrouter API documentation and pricing page. Build the adapter once, keep credentials server-side, and leave room to change routes as models and prices evolve.