Using prxy.monster with Instructor

Instructor (and instructor-js) is a structured-output library that wraps the OpenAI / Anthropic clients. Because Instructor patches existing clients rather than reimplementing them, the integration is the same as the underlying SDK — set the base URL on the client you pass in.

Install

pip install instructor openai pydantic

npm install @instructor-ai/instructor openai zod

Configure

import instructor
from openai import OpenAI
from pydantic import BaseModel
 
client = instructor.from_openai(
    OpenAI(
        base_url="https://api.prxy.monster/v1",
        api_key="prxy_live_xxxxxxxxxxxxxxxxxxxxxxxx",
    )
)
 
class User(BaseModel):
    name: str
    age: int
 
user = client.chat.completions.create(
    model="gpt-4o",
    response_model=User,
    messages=[{"role": "user", "content": "John is 30 years old"}],
)

import Instructor from '@instructor-ai/instructor';
import OpenAI from 'openai';
import { z } from 'zod';
 
const oai = new OpenAI({
  baseURL: 'https://api.prxy.monster/v1',
  apiKey: process.env.OPENAI_API_KEY,
});
 
const client = Instructor({ client: oai, mode: 'TOOLS' });
 
const UserSchema = z.object({
  name: z.string(),
  age: z.number(),
});
 
const user = await client.chat.completions.create({
  model: 'gpt-4o',
  response_model: { schema: UserSchema, name: 'User' },
  messages: [{ role: 'user', content: 'John is 30 years old' }],
});

Code change

The base_url / baseURL arg on the underlying client is the only diff.

You can also use env vars (OPENAI_BASE_URL=https://api.prxy.monster/v1) and pass OpenAI() with no args.

Verify

curl https://api.prxy.monster/health

What you get — extra value for Instructor users

• Semantic cache wins big here — extracting the same schema from similar inputs (invoice line items, contact records, log entries) hits cache often.
• Cost guards prevent runaway extraction loops on bad data.
• Pattern memory — extraction patterns (“when input looks like X, schema fills like Y”) get learned.

Recommended pipeline

For batch extraction:

PRXY_PIPE=exact-cache,semantic-cache,cost-guard,patterns

For interactive extraction:

PRXY_PIPE=semantic-cache,patterns,cost-guard

Anthropic-backed Instructor

instructor.from_anthropic(Anthropic(base_url='https://api.prxy.monster', ...)) works the same way.

Common issues

• Retry loops on validation failure — Instructor will retry the LLM call up to N times if the response doesn’t validate. Each retry counts against your prxy.monster cost-guard budget if enabled. Cap retries at 2-3 for production.
• Streaming structured output (partial_response_model) — pass-through.

Full example

Adapt examples/openai-quickstart — wrap the OpenAI() client in instructor.from_openai(...) and add a Pydantic schema.