python-ai-ml

Python AI/ML Implementation Patterns

Use this skill for LLM-integrated systems, RAG architectures, and AI-adjacent data pipelines. The bias is toward correctness of data flow, cost awareness, and reproducibility — not toward ML research patterns.

Delivery Workflow

AI/ML progress:
- [ ] Step 1: Discover existing pipeline structure, models in use, and vector store schema
- [ ] Step 2: Define data shapes for chunking, embedding, and retrieval contracts
- [ ] Step 3: Implement with explicit model versioning and token budgeting
- [ ] Step 4: Add error handling for API failures, rate limits, and empty retrievals
- [ ] Step 5: Verify retrieval quality, cost controls, and pipeline observability

LLM API Integration

• Pin the model name as a configuration value, never as a string literal in business logic. Model behavior changes between versions.
• Always set max_tokens explicitly — never accept provider defaults that can silently inflate cost.
• Use temperature=0 for deterministic tasks (classification, extraction, code generation); use higher values only when diversity is intentional.
• Implement retry logic with exponential backoff for rate limit and transient errors — use tenacity or equivalent.
• Log token usage per call for cost tracking; surface as metrics in production pipelines.
• Never send raw user input directly to an LLM API in multi-tenant contexts — sanitize or scope prompt injection risks explicitly.

from openai import AsyncOpenAI
from tenacity import retry, stop_after_attempt, wait_exponential

client = AsyncOpenAI()

@retry(stop=stop_after_attempt(3), wait=wait_exponential(multiplier=1, min=2, max=10))
async def complete(
    prompt: str,
    model: str,
    max_tokens: int,
    temperature: float = 0.0,
) -> str:
    response = await client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
        max_tokens=max_tokens,
        temperature=temperature,
    )
    return response.choices[0].message.content or ""