migrate-to-sdk

name	migrate-to-sdk
description	Migrates existing AI code (raw OpenAI/Anthropic SDK calls, LangChain chains, LlamaIndex pipelines) to django-ai-sdk patterns. Produces side-by-side before/after code with a migration checklist. Invoke when the user says "migrate from OpenAI SDK", "migrate from LangChain", "convert my existing agent", "replace openai.ChatCompletion", or "port to django-ai-sdk".
triggers	["migrate from OpenAI SDK","migrate from LangChain","convert my existing agent","replace openai.ChatCompletion","port to django-ai-sdk","convert to django-ai-sdk","migrate from Anthropic SDK","migrate from LlamaIndex","replace direct API calls","convert OpenAI code","migrate existing AI code"]

Migrate Existing AI Code to django-ai-sdk

You are migrating existing AI code to django-ai-sdk. If the user pastes a screenshot or code snippet, analyze it and produce a side-by-side migration.

Step 1 — Identify the Source Pattern

Detect which pattern the existing code uses:

Source pattern	Migration target
openai.chat.completions.create()	Agent.handle()
anthropic.messages.create()	Agent.handle() with provider="anthropic"
LangChain ChatOpenAI + LLMChain	Agent subclass
LangChain ConversationChain + memory	Agent with enable_conversation=True
LangChain tool / agent executor	Agent with tools = [...]
LlamaIndex QueryEngine	Agent with SemanticMemory or RAG tool
Raw requests.post to OpenAI API	Agent.handle()

---

Migration Map: Raw OpenAI → django-ai-sdk

Before (raw openai SDK):

from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

response = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": prompt},
    ],
    temperature=0.7,
    max_tokens=1024,
)
text = response.choices[0].message.content

After (django-ai-sdk):

# myapp/agents.py
from djangosdk.agents.base import Agent


class AssistantAgent(Agent):
    provider = "openai"
    model = "gpt-4.1"
    system_prompt = "You are a helpful assistant."
    temperature = 0.7
    max_tokens = 1024


# Usage:
agent = AssistantAgent()
response = agent.handle(prompt)
text = response.text

What's removed: manual API key passing, manual message list construction, manual response parsing.

---

Migration Map: LangChain LLMChain → django-ai-sdk

Before (LangChain):

from langchain_openai import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain.chains import LLMChain

llm = ChatOpenAI(model="gpt-4o", temperature=0.5)
prompt = ChatPromptTemplate.from_template("Summarize: {text}")
chain = LLMChain(llm=llm, prompt=prompt)

result = chain.invoke({"text": article_text})
summary = result["text"]

After (django-ai-sdk):

from djangosdk.agents.base import Agent


class SummaryAgent(Agent):
    model = "gpt-4.1"
    system_prompt = "You are a summarization expert."
    temperature = 0.5


agent = SummaryAgent()
response = agent.handle(f"Summarize: {article_text}")
summary = response.text

---

Migration Map: LangChain ConversationChain → django-ai-sdk

Before (LangChain + memory):

from langchain_openai import ChatOpenAI
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory

llm = ChatOpenAI(model="gpt-4o")
memory = ConversationBufferMemory()
chain = ConversationChain(llm=llm, memory=memory)

chain.predict(input="Hello!")
chain.predict(input="What did I just say?")

After (django-ai-sdk):

from djangosdk.agents.base import Agent


class ChatAgent(Agent):
    model = "gpt-4.1"
    system_prompt = "You are a helpful assistant."
    enable_conversation = True   # persists history in Django ORM


agent = ChatAgent()
response1 = agent.handle("Hello!", conversation_id="session-001")
response2 = agent.handle("What did I just say?", conversation_id="session-001")

What's different: history is persisted in the database across requests, not in-memory. Pass the same conversation_id from the client.

---

Migration Map: LangChain Agent Executor → django-ai-sdk

Before (LangChain tool-use):

from langchain_openai import ChatOpenAI
from langchain.agents import create_openai_functions_agent, AgentExecutor
from langchain.tools import tool

@tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"It's sunny in {city}"

llm = ChatOpenAI(model="gpt-4o")
agent = create_openai_functions_agent(llm, [get_weather], prompt)
executor = AgentExecutor(agent=agent, tools=[get_weather])
result = executor.invoke({"input": "What's the weather in Istanbul?"})

After (django-ai-sdk):

from djangosdk.agents.base import Agent
from djangosdk.tools.decorator import tool


@tool
def get_weather(city: str) -> str:
    """Get weather for a city."""
    return f"It's sunny in {city}"


class WeatherAgent(Agent):
    model = "gpt-4.1"
    system_prompt = "You are a helpful weather assistant."
    tools = [get_weather]


agent = WeatherAgent()
response = agent.handle("What's the weather in Istanbul?")
print(response.text)

What's removed: AgentExecutor, manual prompt construction, framework-specific @tool import. The dispatch loop is built into Agent.handle().

---

Migration Checklist

After migrating, verify each item:

• Replace openai.OpenAI() / anthropic.Anthropic() client instances with an Agent subclass
• Move api_key to AI_SDK["PROVIDERS"] in settings — remove from code
• Move system / system_prompt to Agent.system_prompt class attribute
• Move model to Agent.model class attribute
• Replace raw @tool (LangChain) with djangosdk.tools.decorator.tool
• Replace ConversationBufferMemory with enable_conversation = True + conversation_id kwarg
• Replace LLMChain pipelines with pipeline(agent_a, agent_b, ...) from orchestration
• Replace AgentExecutor with Agent + tools = [...]
• Replace manual streaming loops with agent.stream() or StreamingChatAPIView
• Replace manual structured output parsing with output_schema = MyPydanticModel
• Add FakeProvider tests — remove any tests that call real APIs

What You Keep

• Custom business logic inside tool functions — just add @tool and type hints
• Pydantic models for structured output — attach as output_schema
• System prompts — paste directly into system_prompt
• Any prompt templates — inline them in handle() calls or in system_prompt