Welcome to FunctAI. This library reimagines how Python developers integrate Large Language Models (LLMs) into their applications. FunctAI allows you to treat AI models as reliable, typed Python functions, abstracting away the complexities of prompt engineering and output parsing.
The core philosophy of FunctAI is simple yet powerful:
The function definition is the prompt, and the function body is the program definition.
By leveraging Python’s native features—docstrings for instructions, type hints for structure, and variable assignments for logic flow—you can define sophisticated AI behaviors with minimal boilerplate.
FunctAI is built on the powerful DSPy framework, unlocking advanced strategies like Chain-of-Thought, automatic optimization, and agentic tool usage through an ergonomic, decorator-based API.
Install FunctAI and its core dependency, DSPy.
pip install functai
Before using FunctAI, you must configure a default Language Model (LM). This requires initializing a DSPy LM provider.
from functai import configure
# Configure FunctAI globally
configure(lm="gpt-4.1", temperature=1.0, api_key="<YOUR_API_KEY>")
Creating an AI function is as simple as defining a standard Python
function with type hints and a docstring, then decorating it with @ai.
from functai import ai, _ai
@ai
def summarize(text: str, focus: str = "key points") -> str:
"""Summarize the text in one concise sentence,
concentrating on the specified focus area."""
# The _ai sentinel represents the LLM output
return _ai
# Call it exactly like a normal Python function
long_text = "FunctAI bridges the gap between Python's expressive syntax and the dynamic capabilities of LLMs. It allows developers to focus on logic rather than boilerplate."
summarize(long_text, focus="developer benefits")
"FunctAI enables developers to concentrate on core logic by reducing boilerplate and leveraging Python's syntax with LLM capabilities."
What happens when you call summarize?
text,
focus).str).@ai DecoratorThe @ai decorator is the magic wand. It transforms a Python function
into an LLM-powered program. It analyzes the function’s signature
(parameters, return type, and docstring) to understand the task
requirements and a prompt is automatically constructed out of that for
you.
@ai
def sentiment(text: str) -> str:
"""Analyze the sentiment of the given text.
Return 'positive', 'negative', or 'neutral'."""
... # An empty body or Ellipsis also works like returning _ai
_ai SentinelThe _ai object is a special sentinel used within an @ai function. It
represents the value(s) that will be generated by the AI. It acts as a
proxy, deferring the actual LM execution.
Returning _ai directly indicates that the LLM’s output is the
function’s final result.
@ai
def extract_price(description: str) -> float:
"""Extract the price from a product description."""
return _ai
FunctAI encourages writing robust code. You can assign _ai to a
variable and apply standard Python operations before returning. _ai
behaves dynamically as if it were the expected return type.
This allows you to combine the power of AI with the reliability of code for validation, cleaning, or transformation.
@ai
def sentiment_score(text: str) -> float:
"""Returns a sentiment score between 0.0 (negative) and 1.0 (positive)."""
# _ai behaves like a float here due to the return type hint
score = _ai
# Post-processing: ensure the score is strictly within bounds
return max(0.0, min(1.0, float(score)))
sentiment_score("I think that FunctAI is amazing!")
0.95
FunctAI excels at extracting structured data. Python type hints serve as the contract between your code and the LLM.
FunctAI handles standard Python types (int, float, bool, str,
list, dict).
@ai
def calculate(expression) -> int:
"""Evaluate the mathematical expression."""
return _ai
result = calculate("What is 15 times 23?")
print(result)
345
@ai
def get_keywords(article) -> list[str]:
"""Extract 5 key terms from the article."""
keywords: list[str] = _ai
# Post-processing example: ensure lowercase
return [k.lower() for k in keywords]
get_keywords("FunctAI excels at extracting structured data. Python type hints serve as the contract between your code and the LLM.")
['functai', 'structured data', 'python type hints', 'contract', 'llm']
For complex data extraction, define a dataclass (or Pydantic model)
and use it as the return type.
from dataclasses import dataclass
from typing import List
@dataclass
class ProductInfo:
name: str
price: float
features: List[str]
in_stock: bool
@ai
def extract_product(description: str) -> ProductInfo:
"""Extract product information from the description."""
return _ai
info = extract_product("iPhone 15 Pro - $999, 5G, titanium design, available now")
# The output is a validated ProductInfo instance
print(info)
ProductInfo(name='iPhone 15 Pro', price=999.0, features=['5G', 'titanium design'], in_stock=True)
Use Enum to restrict the LM’s output to a predefined set of values,
increasing reliability for classification tasks.
from enum import Enum
class TicketPriority(Enum):
LOW = "low"
MEDIUM = "medium"
HIGH = "high"
@ai
def classify_priority(issue_description: str) -> TicketPriority:
"""Analyzes the issue and classifies its priority level."""
return _ai
result = classify_priority(issue_description="The main database is unresponsive.")
result
<TicketPriority.HIGH: 'high'>
You can also use typing Literal:
from typing import Literal
@ai
def categorize(text: str) -> Literal["sport", "fashion"]:
...
categorize("Vibe Coding as been declared a sport")
'sport'
FunctAI uses a flexible, cascading configuration system. Settings are applied in the following order of precedence (highest to lowest):
@ai(temperature=0.1))with configure(temperature=0.1):)configure(temperature=0.1))Use functai.configure() for project-wide defaults.
import functai
# Example using GPT-3.5-Turbo as a default
functai.configure(
lm="groq/openai/gpt-oss-120b",
temperature=0.5
)
@ai
def math(formal):
...
math("2+2")
2025/08/30 13:32:38 WARNING dspy.adapters.json_adapter: Failed to use structured output format, falling back to JSON mode.
4
Override defaults for specific functions directly in the decorator. This is useful when different tasks require different models or creativity levels.
# Deterministic task requiring a powerful model
@ai(temperature=0.0, lm="openai/gpt-4o")
def legal_analysis(document):
"""Provide precise legal analysis of the document."""
return _ai
# Creative task using a different provider
@ai(temperature=0.9, lm="anthropic/claude-sonnet-4-20250514")
def creative_story(prompt):
"""Write a creative story based on the prompt."""
return _ai
Use the functai.configure() context manager to temporarily override
defaults for a block of code.
from functai import configure
@ai
def analyze(data): return _ai
analyze("data1") # Uses global defaults
# Temporarily switch model and temperature
with configure(temperature=0.0, lm="gpt-4o"):
print(analyze("data2")) # Uses GPT-4.1, Temp 0.0
analyze("data3") # Back to global defaults
2025/08/30 13:32:43 WARNING dspy.adapters.json_adapter: Failed to use structured output format, falling back to JSON mode.
2025/08/30 13:32:49 WARNING dspy.adapters.json_adapter: Failed to use structured output format, falling back to JSON mode.
data2
'data3'
FunctAI truly shines by allowing developers to define complex execution strategies directly within the function body, adhering to the “function body is the program definition” philosophy.
Eliciting step-by-step reasoning (Chain of Thought) often significantly improves the quality and accuracy of the final answer, especially for complex tasks.
In FunctAI, you define CoT by declaring intermediate reasoning steps
within the function body using _ai assignments with descriptions.
@ai
def solve_math_problem(question: str) -> float:
"""Solves a math word problem and returns the numerical answer."""
# Define the reasoning step:
# 1. The variable name ('reasoning') becomes the field name.
# 2. The type hint (str) defines the output type for this step.
# 3. The subscript _ai["..."] provides specific instructions for the LLM.
reasoning: str = _ai["Step-by-step thinking process to reach the solution."]
# The final return value (float) is the main output
return _ai
Behavior: FunctAI analyzes the function body, detects the
intermediate reasoning assignment, and automatically configures the
execution to generate the reasoning before attempting to generate the
final result.
(Note: You can also enable a generic CoT by setting
@ai(module="cot"), but the explicit definition above offers more
control.)
While the function call normally returns only the final result, you can
access the intermediate steps (like reasoning) by adding the special
argument all=True to the function call.
This returns the raw prediction object containing all generated fields.
question = "If a train travels 120 miles in 2 hours, what is its speed?"
prediction = solve_math_problem(question, all=True)
print("--- Reasoning ---")
print(prediction.reasoning)
print("\n--- Answer ---")
# When _ai is returned directly, the main output is stored in the 'result' attribute
print(prediction.result)
--- Reasoning ---
Speed is calculated by dividing distance by time. The train travels 120 miles in 2 hours, so speed = 120 miles / 2 hours = 60 miles per hour.
--- Answer ---
60.0
You can define and return multiple distinct outputs from a single function call by declaring them inline, similar to the CoT pattern.
from typing import Tuple
@ai()
def critique_and_improve(text: str) -> Tuple[str, str, int]:
"""
Analyze the text, provide constructive criticism, and suggest an improved version.
"""
# Define explicit output fields using _ai[...]
critique: str = _ai["Constructive criticism focusing on clarity and tone."]
improved_text: str = _ai["The improved version of the text."]
# Return the materialized fields (Python handles the Tuple structure)
return critique, improved_text, 1.0
critique, improved, number = critique_and_improve(text="U should fix this asap, it's broken.")
print("--- Critique ---")
print(critique)
print("--- Improved ---")
print(improved)
--- Critique ---
The original message is overly informal and uses shorthand ('U') and vague urgency ('asap') that may come across as unprofessional. It lacks specific details about what is broken, which can make it harder for the recipient to address the issue efficiently. A clearer, more courteous tone with a brief description of the problem would improve communication.
--- Improved ---
Please address this issue as soon as possible; the current functionality appears to be broken.
FunctAI supports the ReAct (Reasoning + Acting) pattern for creating agents that can interact with external tools. Tools are standard, typed Python functions.
When the tools argument is provided to the @ai decorator, the
execution strategy automatically upgrades to an agentic loop (using
dspy.ReAct).
# 1. Define tools
def search_web(query: str) -> str:
"""Searches the web for information. (Mock implementation)"""
print(f"[Tool executing: Searching for '{query}']")
# In a real scenario, this would call a search API
return f"Mock search results for {query}."
def calculate(expression: str) -> float:
"""Performs mathematical calculations. (Mock implementation)"""
print(f"[Tool executing: Calculating '{expression}']")
# WARNING: eval() is unsafe in production. Use a safe math library.
return eval(expression)
# 2. Define the AI function with access to the tools
@ai(tools=[search_web, calculate])
def research_assistant(question: str) -> str:
"""Answer questions using available tools to gather data and perform calculations."""
return _ai
# 3. Execute the agent
# The AI will potentially use search_web and then calculate.
answer = research_assistant("What is the result of (15 * 23) + 10?")
[Tool executing: Calculating '(15 * 23) + 10']
2025/08/30 13:32:58 WARNING dspy.adapters.json_adapter: Failed to use structured output format, falling back to JSON mode.
import functai
#| echo: false
print(functai.phistory())
[2025-08-30T13:32:58.957207]
System message:
Your input fields are:
1. `question` (str):
2. `trajectory` (str):
Your output fields are:
1. `reasoning` (str):
2. `result` (str):
All interactions will be structured in the following way, with the appropriate values filled in.
Inputs will have the following structure:
[[ ## question ## ]]
{question}
[[ ## trajectory ## ]]
{trajectory}
Outputs will be a JSON object with the following fields.
{
"reasoning": "{reasoning}",
"result": "{result}"
}
In adhering to this structure, your objective is:
Function: research_assistant
Answer questions using available tools to gather data and perform calculations.
User message:
[[ ## question ## ]]
What is the result of (15 * 23) + 10?
[[ ## trajectory ## ]]
[[ ## thought_0 ## ]]
I need to compute the arithmetic expression (15 * 23) + 10.
[[ ## tool_name_0 ## ]]
calculate
[[ ## tool_args_0 ## ]]
{"expression": "(15 * 23) + 10"}
[[ ## observation_0 ## ]]
355
[[ ## thought_1 ## ]]
The calculation is complete; the result of (15 * 23) + 10 is 355.
[[ ## tool_name_1 ## ]]
finish
[[ ## tool_args_1 ## ]]
{}
[[ ## observation_1 ## ]]
Completed.
Respond with a JSON object in the following order of fields: `reasoning`, then `result`.
Response:
{"reasoning":"I used the calculate tool to evaluate the expression (15 * 23) + 10, which gave 355. No further steps are needed.","result":"355"}
Behavior: The function will iteratively think about the task, decide which tool to use, execute the tool, observe the results, and repeat until it can provide the final answer.
By default, @ai functions are stateless; each call is independent. To
maintain context across calls (e.g., in a chatbot scenario), set
stateful=True.
from functai import ai, _ai
@ai(lm="gpt-4.1", stateful=True)
def assistant(message):
"""A friendly AI assistant that remembers the conversation history."""
return _ai
response1 = assistant("Hello, my name is Alex.")
print(f"Output1: {response1}")
response2 = assistant("What is my name?")
print(f"Output2: {response2}")
Output1: Hello Alex! It's nice to meet you. How can I assist you today?
Output2: Your name is Alex.
[2025-08-30T13:32:59.675591]
System message:
Your input fields are:
1. `message` (str):
2. `history` (History):
Your output fields are:
1. `result` (Any):
All interactions will be structured in the following way, with the appropriate values filled in.
Inputs will have the following structure:
[[ ## message ## ]]
{message}
[[ ## history ## ]]
{history}
Outputs will be a JSON object with the following fields.
{
"result": "{result} # note: the value you produce must adhere to the JSON schema: {}"
}
In adhering to this structure, your objective is:
Function: assistant
A friendly AI assistant that remembers the conversation history.
User message:
[[ ## message ## ]]
Hello, my name is Alex.
Assistant message:
{
"result": "Hello Alex! It's nice to meet you. How can I assist you today?"
}
User message:
[[ ## message ## ]]
What is my name?
Respond with a JSON object in the following order of fields: `result` (must be formatted as a valid Python Any).
Response:
{
"result": "Your name is Alex."
}
Behavior: When stateful=True, FunctAI automatically includes the
history of previous inputs and outputs in the context of the next call.
WARNING: optimization is still a work in progress
FunctAI integrates seamlessly with DSPy’s optimization capabilities (Teleprompters). Optimization (often called compilation in DSPy) improves the quality and reliability of your AI functions by using a dataset of examples.
The optimizer can automatically generate effective few-shot examples or
refine instructions. This happens in place using the .opt() method
on the function object.
import dspy
from dspy import Example
from functai import ai, _ai
dspy.configure(lm = dspy.LM("gpt-4.1"))
# 1. Define the function
@ai
def classify_intent(user_query: str) -> str:
"""Classify user intent as 'booking', 'cancelation', or 'information'."""
return _ai
# 2. Define the training data (List of DSPy Examples)
# .with_inputs() specifies which keys are inputs to the function
trainset = [
Example(user_query="I need to reserve a room.", result="booking").with_inputs("user_query"),
Example(user_query="How do I get there?", result="information").with_inputs("user_query"),
Example(user_query="I want to cancel my reservation.", result="cancelation").with_inputs("user_query"),
]
# 3. Optimize the function in place
# strategy="launch" typically uses a default like BootstrapFewShot
print("Optimizing...")
classify_intent.opt(trainset=trainset)
print("Optimization complete.")
# 4. The function is now optimized (it includes generated few-shot examples in its prompt)
result = classify_intent("Can I book a suite for next Tuesday?")
# Output: "booking"
Optimizing...
0%| | 0/3 [00:00<?, ?it/s]100%|██████████| 3/3 [00:00<00:00, 36.98it/s]
Bootstrapped 3 full traces after 2 examples for up to 1 rounds, amounting to 3 attempts.
Optimization complete.
2025/08/30 13:33:04 WARNING dspy.adapters.json_adapter: Failed to use structured output format, falling back to JSON mode.
FunctAI tracks optimization steps. If the results are not satisfactory,
you can revert using .undo_opt().
# Revert the last optimization step
classify_intent.undo_opt(steps=1)
More to come
To see the last call, use functai.phistory().
from functai import phistory
# After running some AI functions...
print(phistory()) # Show the last call
[2025-08-30T13:33:06.742620]
System message:
Your input fields are:
1. `user_query` (str):
Your output fields are:
1. `result` (str):
All interactions will be structured in the following way, with the appropriate values filled in.
Inputs will have the following structure:
[[ ## user_query ## ]]
{user_query}
Outputs will be a JSON object with the following fields.
{
"result": "{result}"
}
In adhering to this structure, your objective is:
Function: classify_intent
Classify user intent as 'booking', 'cancelation', or 'information'.
User message:
[[ ## user_query ## ]]
I need to reserve a room.
Assistant message:
{
"result": "booking"
}
User message:
[[ ## user_query ## ]]
How do I get there?
Assistant message:
{
"result": "information"
}
User message:
[[ ## user_query ## ]]
I want to cancel my reservation.
Assistant message:
{
"result": "cancelation"
}
User message:
[[ ## user_query ## ]]
Can I book a suite for next Tuesday?
Respond with a JSON object in the following order of fields: `result`.
Response:
{
"result": "booking"
}
This example demonstrates chaining multiple AI functions to process unstructured data reliably.
from dataclasses import dataclass
from typing import List
from functai import ai, _ai, configure
# Ensure deterministic extraction
configure(lm = "gpt-4.1", temperature=0.0, adapter="json")
# 1. Define the target structure
@dataclass
class Invoice:
invoice_number: str
vendor_name: str
total: float
items: List[str]
# 2. Define the extraction function with CoT for accuracy
@ai
def extract_invoice(document_text: str) -> Invoice:
"""Extract invoice information from the document text.
Parse all relevant fields accurately. Convert amounts to float.
"""
thought_process: str = _ai["Analyze the document layout and identify the location of each field before extracting."]
return _ai
# 3. Define a validation function
@ai
def validate_invoice(invoice: Invoice) -> bool:
"""Validate if the invoice data is complete and reasonable.
Check if the total is positive and required fields are present.
"""
return _ai
# 4. Define a summarization function
@ai
def summarize_invoice(invoice: Invoice) -> str:
"""Create a brief, human-readable summary of the invoice."""
return _ai
# 5. Execute the pipeline
document = """
INVOICE
Vendor: TechCorp Inc.
Invoice #: INV-2025-101
Items: 5x Laptops, 2x Monitors
Total: $5600.00
"""
invoice = extract_invoice(document)
if validate_invoice(invoice):
summary = summarize_invoice(invoice)
print("Invoice Validated Successfully!")
print(summary)
else:
print("Invoice Validation Failed.")
Invoice Validated Successfully!
Invoice INV-2025-101 from TechCorp Inc. totals $5,600.00 and includes 5 laptops and 2 monitors.
This example builds a sophisticated agent using tools and structured internal outputs.
from functai import ai, _ai, configure
configure(lm = "gpt-4.1")
# Define Tools (Placeholders)
def search_web(query: str) -> str:
"""Search the web for information."""
print(f"[Searching: {query}]")
return f"Mock search results for {query}."
def read_paper(paper_id: str) -> str:
"""Read the content of a specific research paper."""
print(f"[Reading: {paper_id}]")
return f"Mock content of paper {paper_id}."
# Define the Agent
@ai(tools=[search_web, read_paper])
def research_assistant(query: str) -> str:
"""Advanced research assistant.
Use available tools to gather information. Synthesize findings.
"""
# Define intermediate outputs for better structure and inspection
research_notes: list[str] = _ai["Key findings gathered during the ReAct process."]
confidence: str = _ai["Confidence level: high/medium/low."]
sources: list[str] = _ai["Sources consulted during the ReAct process."]
answer = _ai
# Post-processing: Add metadata to the final response
return f"{answer}\n\nConfidence: {confidence}\nSources: {', '.join(sources)}"
# Execution
research_assistant("What are the latest breakthroughs in quantum computing?")
[Searching: latest breakthroughs in quantum computing 2024]
"The latest breakthroughs in quantum computing as of 2024 include IBM's unveiling of the 1,121-qubit 'Condor' processor, which represents a major step forward in hardware scalability. Researchers at Google, Microsoft, and Quantinuum have achieved significant progress in quantum error correction, with logical qubits now outperforming physical qubits in some cases. Advances in quantum networking have enabled entanglement distribution over longer distances, moving closer to a functional quantum internet. Additionally, new quantum algorithms and hybrid approaches are being developed for practical applications in chemistry, optimization, and machine learning. Commercial access to quantum computing continues to grow, with more cloud-based services available to researchers and businesses.\n\nConfidence: high\nSources: https://www.ibm.com/blog/quantum-condor-1121-qubit-processor/, https://www.nature.com/articles/d41586-024-00000-0, https://www.quantamagazine.org/quantum-error-correction-breakthroughs-2024-20240110/, https://www.microsoft.com/en-us/research/blog/advances-in-fault-tolerant-quantum-computing-2024/, https://www.scientificamerican.com/article/quantum-internet-milestones-2024/"
FAQs
DSPy-powered function decorators for AI-enhanced programming
We found that functai demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Did you know?
Socket for GitHub automatically highlights issues in each pull request and monitors the health of all your open source dependencies. Discover the contents of your packages and block harmful activity before you install or update your dependencies.
Security News
A data handling bug in OSV.dev caused disputed CVEs to disappear from vulnerability feeds until a recent fix restored over 500 advisories.
Research
/Security News
175 malicious npm packages (26k+ downloads) used unpkg CDN to host redirect scripts for a credential-phishing campaign targeting 135+ organizations worldwide.
Security News
Python 3.14 adds template strings, deferred annotations, and subinterpreters, plus free-threaded mode, an experimental JIT, and Sigstore verification.