Delegate Routine Work to Cheaper Models

from dataclasses import dataclass
from enum import Enum

class Tier(Enum):
    FREE     = "free"
    ECONOMY  = "economy"
    MID      = "mid"
    PREMIUM  = "premium"

@dataclass
class Task:
    name: str
    prompt: str
    tier: Tier
    max_tokens: int = 512

TIER_MODELS = {
    Tier.FREE:    "meta-llama/llama-4-maverick",  # swap for a free model like cohere/north-mini-code:free
    Tier.ECONOMY: "meta-llama/llama-4-maverick",
    Tier.MID:     "anthropic/claude-sonnet-5",
    Tier.PREMIUM: "anthropic/claude-opus-4.8",
}

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://opendunes.com/api/v1",
    api_key=os.environ["OPENDUNES_API_KEY"],
)

def dispatch(task: Task, context: str = "") -> str:
    model = TIER_MODELS[task.tier]
    messages = []
    if context:
        messages.append({"role": "user", "content": f"Context:\n{context}"})
    messages.append({"role": "user", "content": task.prompt})

    resp = client.chat.completions.create(
        model=model,
        messages=messages,
        max_tokens=task.max_tokens,
        temperature=0 if task.tier in (Tier.FREE, Tier.ECONOMY) else 0.7,
    )
    return resp.choices[0].message.content or ""

def process_feedback(feedback: str) -> dict:
    # Step 1: classify — easy, use economy tier
    sentiment = dispatch(Task(
        name="classify_sentiment",
        prompt=f"Classify the sentiment of this feedback as positive, neutral, or negative. Respond with one word only.\n\n{feedback}",
        tier=Tier.ECONOMY,
        max_tokens=5,
    ))

    # Step 2: extract themes — moderate, use economy tier
    themes = dispatch(Task(
        name="extract_themes",
        prompt=f"List up to 3 key themes in this feedback, one per line, no bullets.\n\n{feedback}",
        tier=Tier.ECONOMY,
        max_tokens=100,
    ))

    # Step 3: draft a response — nuanced, use mid tier
    draft = dispatch(Task(
        name="draft_response",
        prompt=f"Write a polite, helpful one-paragraph response to this customer feedback.\n\nFeedback: {feedback}\nSentiment: {sentiment}\nThemes: {themes}",
        tier=Tier.MID,
        max_tokens=200,
    ))

    # Step 4: final quality check on negative feedback only — expensive, use premium
    if sentiment.strip().lower() == "negative":
        draft = dispatch(Task(
            name="refine_negative_response",
            prompt=f"Improve this draft response to a negative customer complaint. Be empathetic and concrete.\n\nDraft: {draft}\n\nOriginal feedback: {feedback}",
            tier=Tier.PREMIUM,
            max_tokens=300,
        ))

    return {"sentiment": sentiment, "themes": themes, "response_draft": draft}

def infer_tier(prompt: str) -> Tier:
    """Rough heuristic: long prompts and certain keywords signal harder tasks."""
    word_count = len(prompt.split())
    hard_keywords = {"reason", "analyze", "compare", "design", "debug", "refactor", "evaluate"}
    if word_count > 300 or any(k in prompt.lower() for k in hard_keywords):
        return Tier.MID
    if word_count > 600 or "complex" in prompt.lower():
        return Tier.PREMIUM
    return Tier.ECONOMY