Python FastAPI + LLM Backend

Building AI-powered Python backends — the patterns that make them reliable, fast, and maintainable.

Project Structure

backend/
├── main.py                  # FastAPI app factory
├── routers/
│   ├── generate.py          # AI generation endpoints
│   ├── jobs.py              # Background job management
│   └── health.py            # Health + readiness checks
├── services/
│   ├── llm.py               # LLM client abstraction
│   ├── generator.py         # Business logic for generation
│   └── storage.py           # File / DB persistence
├── models/
│   ├── requests.py          # Pydantic input models
│   └── responses.py         # Pydantic output models
├── core/
│   ├── config.py            # Settings (pydantic-settings)
│   ├── logging.py           # Structured logging
│   └── exceptions.py        # Custom exceptions + handlers
├── tests/
└── requirements.txt

FastAPI App Factory

# main.py
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from contextlib import asynccontextmanager
from core.config import settings
from routers import generate, jobs, health

@asynccontextmanager
async def lifespan(app: FastAPI):
    # startup: initialize clients, warm caches
    yield
    # shutdown: close connections

app = FastAPI(title="AI Backend", lifespan=lifespan)

app.add_middleware(
    CORSMiddleware,
    allow_origins=settings.allowed_origins,
    allow_methods=["*"],
    allow_headers=["*"],
)

app.include_router(generate.router, prefix="/api/v1/generate", tags=["generate"])
app.include_router(jobs.router, prefix="/api/v1/jobs", tags=["jobs"])
app.include_router(health.router, tags=["health"])

Settings (pydantic-settings)

# core/config.py
from pydantic_settings import BaseSettings
from functools import lru_cache

class Settings(BaseSettings):
    # LLM
    anthropic_api_key: str = ""
    openai_api_key: str = ""
    default_model: str = "claude-sonnet-4-6"
    max_tokens: int = 8096

    # App
    allowed_origins: list[str] = ["http://localhost:5173"]
    debug: bool = False

    class Config:
        env_file = ".env"
        case_sensitive = False

@lru_cache
def get_settings() -> Settings:
    return Settings()

settings = get_settings()

LLM Client — LiteLLM (recommended for multi-provider)

LiteLLM gives one unified API across Anthropic, OpenAI, Azure, Bedrock, etc.

# services/llm.py
import litellm
from litellm import acompletion, completion
from core.config import settings

litellm.drop_params = True  # ignore unsupported params per provider

async def call_llm(
    messages: list[dict],
    model: str | None = None,
    response_format: type | None = None,  # Pydantic model for structured output
    temperature: float = 0.2,
    max_tokens: int | None = None,
) -> str:
    response = await acompletion(
        model=model or settings.default_model,
        messages=messages,
        temperature=temperature,
        max_tokens=max_tokens or settings.max_tokens,
        response_format=response_format,
    )
    return response.choices[0].message.content

Structured Outputs with Pydantic

from pydantic import BaseModel, Field
from typing import Literal

class DbtModel(BaseModel):
    model_name: str = Field(description="snake_case dbt model name")
    materialization: Literal["view", "table", "incremental"] = "view"
    sql: str = Field(description="The full SQL content of the model")
    description: str

class GeneratedProject(BaseModel):
    models: list[DbtModel]
    sources_yaml: str
    schema_yaml: str
    issues: list[str] = Field(default_factory=list)

async def generate_dbt_project(mapping_json: dict) -> GeneratedProject:
    response = await call_llm(
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": f"Convert this mapping:\n{mapping_json}"}
        ],
        response_format=GeneratedProject,  # LiteLLM handles JSON schema injection
    )
    return GeneratedProject.model_validate_json(response)

Anthropic SDK Directly (for advanced features)

import anthropic
from core.config import settings

client = anthropic.AsyncAnthropic(api_key=settings.anthropic_api_key)

async def call_claude(
    system: str,
    user_message: str,
    model: str = "claude-sonnet-4-6",
) -> str:
    message = await client.messages.create(
        model=model,
        max_tokens=8096,
        system=system,
        messages=[{"role": "user", "content": user_message}],
    )
    return message.content[0].text

# Tool use (function calling)
async def call_claude_with_tools(
    messages: list[dict],
    tools: list[dict],  # Anthropic tool schema format
) -> anthropic.types.Message:
    return await client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=8096,
        tools=tools,
        messages=messages,
    )

Streaming Responses (SSE)

# routers/generate.py
from fastapi import APIRouter
from fastapi.responses import StreamingResponse
import anthropic
import json

router = APIRouter()
client = anthropic.AsyncAnthropic()

@router.post("/stream")
async def generate_stream(request: GenerateRequest):
    async def event_generator():
        async with client.messages.stream(
            model="claude-sonnet-4-6",
            max_tokens=8096,
            messages=[{"role": "user", "content": request.prompt}],
        ) as stream:
            async for text in stream.text_stream:
                yield f"data: {json.dumps({'text': text})}\n\n"
        yield "data: [DONE]\n\n"

    return StreamingResponse(event_generator(), media_type="text/event-stream")

File Upload for AI Processing

# routers/generate.py
from fastapi import UploadFile, File, HTTPException
import json

@router.post("/from-mapping")
async def generate_from_mapping(file: UploadFile = File(...)):
    if not file.filename.endswith((".json", ".xml")):
        raise HTTPException(400, "Only .json or .xml mapping files supported")

    content = await file.read()
    try:
        mapping = json.loads(content)
    except json.JSONDecodeError:
        mapping = content.decode("utf-8")  # XML — parse downstream

    result = await generator_service.process(mapping)
    return result

Background Jobs (long-running AI tasks)

# For generation tasks that take > 30s, use background jobs
from fastapi import BackgroundTasks
import uuid
from services.storage import job_store

@router.post("/async-generate")
async def start_generation(request: GenerateRequest, background_tasks: BackgroundTasks):
    job_id = str(uuid.uuid4())
    await job_store.create(job_id, status="pending")
    background_tasks.add_task(run_generation, job_id, request)
    return {"job_id": job_id, "status": "pending"}

@router.get("/jobs/{job_id}")
async def get_job_status(job_id: str):
    job = await job_store.get(job_id)
    if not job:
        raise HTTPException(404, "Job not found")
    return job

async def run_generation(job_id: str, request: GenerateRequest):
    try:
        await job_store.update(job_id, status="running")
        result = await generator_service.process(request)
        await job_store.update(job_id, status="done", result=result)
    except Exception as e:
        await job_store.update(job_id, status="error", error=str(e))

Retry + Error Handling

# services/llm.py
import asyncio
from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type
import litellm

@retry(
    stop=stop_after_attempt(3),
    wait=wait_exponential(multiplier=1, min=2, max=10),
    retry=retry_if_exception_type((litellm.RateLimitError, litellm.APIConnectionError)),
    reraise=True,
)
async def call_llm_with_retry(messages: list[dict], **kwargs) -> str:
    return await call_llm(messages, **kwargs)

# core/exceptions.py
from fastapi import Request, HTTPException
from fastapi.responses import JSONResponse
import litellm

async def llm_exception_handler(request: Request, exc: litellm.APIError):
    return JSONResponse(
        status_code=502,
        content={"error": "LLM provider error", "detail": str(exc)},
    )

# In main.py:
app.add_exception_handler(litellm.APIError, llm_exception_handler)

LLM Monitoring / Structured Logging

# services/llm_monitor.py
import time
from dataclasses import dataclass, field
from datetime import datetime
import json

@dataclass
class LLMCall:
    model: str
    prompt_tokens: int = 0
    completion_tokens: int = 0
    latency_ms: float = 0
    success: bool = True
    error: str | None = None
    timestamp: str = field(default_factory=lambda: datetime.utcnow().isoformat())

class LLMMonitor:
    def __init__(self, job_id: str):
        self.job_id = job_id
        self.calls: list[LLMCall] = []

    async def track(self, coro, model: str) -> str:
        start = time.monotonic()
        try:
            result = await coro
            elapsed = (time.monotonic() - start) * 1000
            # extract token counts if available
            self.calls.append(LLMCall(model=model, latency_ms=elapsed))
            return result
        except Exception as e:
            elapsed = (time.monotonic() - start) * 1000
            self.calls.append(LLMCall(model=model, latency_ms=elapsed, success=False, error=str(e)))
            raise

    def summary(self) -> dict:
        return {
            "job_id": self.job_id,
            "total_calls": len(self.calls),
            "total_latency_ms": sum(c.latency_ms for c in self.calls),
            "success_rate": sum(c.success for c in self.calls) / len(self.calls) if self.calls else 0,
        }

Prompt Engineering Patterns

System prompt structure — be explicit about output format:

SYSTEM_PROMPT = """
You are a dbt model generator. Given an Informatica mapping, produce a valid dbt project.

ALWAYS respond with valid JSON matching this schema:
{schema_json}

Rules:
- Use snake_case for all model names
- Always use ref() and source() — never hardcode schema/table
- Cast all columns to their final types in staging models
"""

Few-shot examples — include in the user message, not the system prompt:

def build_user_message(mapping: dict, examples: list[dict]) -> str:
    parts = []
    for ex in examples[:2]:  # 2 examples is usually enough
        parts.append(f"Input:\n{json.dumps(ex['input'])}\n\nOutput:\n{json.dumps(ex['output'])}")
    parts.append(f"Input:\n{json.dumps(mapping)}\n\nOutput:")
    return "\n\n---\n\n".join(parts)

Chain of thought for complex transformations:

# Add a "thinking" step before the final JSON output
messages = [
    {"role": "user", "content": "Analyze this mapping and convert it..."},
    # Let Claude reason freely first, then constrain the output
]
# Use extended thinking (Anthropic) or separate reasoning step

Testing AI Endpoints

# tests/test_generate.py
import pytest
from httpx import AsyncClient
from main import app

@pytest.mark.asyncio
async def test_generate_from_mapping():
    async with AsyncClient(app=app, base_url="http://test") as client:
        response = await client.post(
            "/api/v1/generate/from-mapping",
            files={"file": ("mapping.json", b'{"sources": [...]}', "application/json")},
        )
    assert response.status_code == 200
    result = response.json()
    assert "models" in result
    assert len(result["models"]) > 0

# Mock LLM calls in tests to avoid costs:
@pytest.fixture(autouse=True)
def mock_llm(monkeypatch):
    async def fake_call_llm(messages, **kwargs):
        return '{"models": [], "sources_yaml": "", "schema_yaml": ""}'
    monkeypatch.setattr("services.llm.call_llm", fake_call_llm)

Docker Setup

# Dockerfile
FROM python:3.12-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
COPY . .
EXPOSE 8000
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]

# docker-compose.yml (dev)
services:
  api:
    build: .
    ports: ["8000:8000"]
    env_file: .env
    volumes:
      - .:/app       # hot-reload in dev
    command: uvicorn main:app --host 0.0.0.0 --port 8000 --reload

Requirements

fastapi>=0.115
uvicorn[standard]>=0.30
pydantic>=2.0
pydantic-settings>=2.0
litellm>=1.40      # multi-provider LLM unified API
anthropic>=0.40    # direct Anthropic SDK
tenacity>=8.0      # retry logic
httpx>=0.27        # async HTTP client
python-multipart   # file uploads