incredible-python 0.1.2

Incredible API Python SDK compatible with Anthropics' client interface

Incredible Python SDK

Drop-in replacement for the Anthropic Python SDK that speaks to the Incredible API. The goal is plug-and-play migration: swap anthropic.Anthropic for incredible_python.Incredible and keep the rest of your code unchanged—function calling, streaming, token counting, and multi-step workflows included.

---

Table of Contents

1. Features
2. Installation
3. Configuration
4. Quick Start
5. Function Calling
6. Streaming Responses
7. Token Counting
8. Cookbook Compatibility
9. Testing & Development
10. Contributing
11. License

---

Features

• Anthropic-compatible client: identical constructor and method signatures where possible.
• Function/tool calling: pass functions (and tools) to messages.create and receive structured tool-use events.
• Streaming support: iterate over server-sent events that follow Anthropic semantics (data: ..., [DONE]).
• Utility helpers: helper functions to create function_call / function_call_result messages when you run tools locally.
• Token counting: call client.messages.count_tokens(...) to estimate usage before executing a request.
• Cookbook-ready: all examples in Incredible-API-Cookbook-main/ can now be executed with this SDK.

---

Installation

From source (recommended during development)

# create and activate a virtual environment (optional but recommended)
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

# install in editable mode
pip install -e .

(Future) From PyPI

When the package is published you will be able to run:

pip install incredible-python

---

Configuration

The client reads configuration from keyword arguments or environment variables—mirroring the Anthropic SDK.

Setting	Env Var	Default
API key	INCREDIBLE_API_KEY	(required)
Base URL	INCREDIBLE_BASE_URL	https://api.incredible.one
Timeout	—	600 seconds
Max retries	—	2 (with backoff)

export INCREDIBLE_API_KEY="your-secret-key"

---

Quick Start

from incredible_python import Incredible

client = Incredible()
response = client.messages.create(
    model="small-1",
    max_tokens=256,
    messages=[
        {"role": "user", "content": "Give me three startup ideas."}
    ],
)
print(response.content[0]["text"])

response is a rich object that exposes Anthropic-style properties:

print(response.tool_calls)     # list of ToolCall objects (if any)
print(response.usage)          # token usage metadata (if provided)
print(response.stop_reason)    # e.g. "end_turn", "stop_sequence"

---

Function Calling

Provide tool schemas via functions

Important: If you need to work with large context (e.g., multi-megabyte datasets or long documents), register a function/tool that returns that data instead of pasting it directly into a message. The Incredible API assumes heavy payloads are fetched on demand via tool calls; injecting them into the prompt will overflow the current context window.

(or tools) when calling messages.create. If the Incredible model decides to call your function, you can execute it locally and send a follow-up message using the helper utilities.

Manual control

from incredible_python import Incredible, helpers

client = Incredible()

functions = [
    {
        "name": "calculate_operation",
        "description": "Perform basic math",
        "parameters": {
            "type": "object",
            "properties": {
                "operation": {"type": "string", "enum": ["add", "subtract", "multiply", "divide"]},
                "a": {"type": "number"},
                "b": {"type": "number"},
            },
            "required": ["operation", "a", "b"],
        },
    }
]

initial = client.messages.create(
    model="small-1",
    max_tokens=64,
    messages=[{"role": "user", "content": "What is 24 multiplied by 7?"}],
    functions=functions,
)

for call in initial.tool_calls:
    result_value = call.arguments["a"] * call.arguments["b"]
    function_result = helpers.make_tool_result(call.id, result_value)

    follow_up_messages = [
        {"role": "user", "content": "What is 24 multiplied by 7?"},
        helpers.make_tool_call(call.name, call.arguments, call_id=call.id),
        function_result,
    ]

    final = client.messages.create(
        model="small-1",
        max_tokens=128,
        messages=follow_up_messages,
        functions=functions,
    )
    print(final.content)

Automated helper flow

When the model returns a function_call block you can offload the parsing and follow-up message creation to the helper utilities:

from incredible_python import Incredible, helpers

client = Incredible()
registry = {
    "calculate_operation": lambda operation, a, b: eval(f"{a}{ {'add':'+','subtract':'-','multiply':'*','divide':'/' }[operation] }{b}"),
}

messages = [{"role": "user", "content": "Calculate 42 / 7."}]
response = client.messages.create(
    model="small-1",
    max_tokens=64,
    messages=messages,
    functions=list_registry_as_json_schema(...),  # supply your schemas as before
)

plan = helpers.build_tool_execution_plan(response.raw)
if plan and not plan.is_empty():
    results = helpers.execute_plan(plan, registry=registry)
    follow_up = helpers.build_follow_up_messages(messages, plan, results)
    final = client.messages.create(
        model="small-1",
        max_tokens=64,
        messages=follow_up,
        functions=list_registry_as_json_schema(...),
    )
    print(final.content)

Tip: execute_plan takes care of running multiple tool calls in the order the model requested. build_follow_up_messages packages the results back into the canonical function_call_result format required by the API.

---

Streaming Responses

client = Incredible()
stream = client.messages.stream(
    model="small-1",
    max_tokens=256,
    messages=[{"role": "user", "content": "Stream a short poem about the ocean."}],
)

for event in stream.iter_lines():
    if event.get("content", {}).get("type") == "content_chunk":
        chunk = event["content"]["content"]
        print(chunk, end="", flush=True)

StreamBuilder.iter_lines() yields parsed JSON events; when the server sends data: [DONE] the iterator stops. Call stream.close() to release the network connection early if needed.

---

Token Counting

usage_estimate = client.messages.count_tokens(
    model="small-1",
    messages=[{"role": "user", "content": "Summarise this 1,000 word article."}],
)
print(usage_estimate)

Note: The live Incredible API has not yet exposed /v1/messages/count_tokens, so this method will raise a 404 until the endpoint ships. The SDK keeps the method for future compatibility and mirrors Anthropic’s signature.

---

Cookbook Compatibility

This SDK was validated against every scenario in Incredible-API-Cookbook-main:

• ✅ 01-getting-started/ – basic chat, conversations, streaming, function calling
• ✅ 02-function-calling/ – multiple tools, advanced multi-step workflows, JSON extraction, and more

To migrate an example:

1. Replace raw requests.post(...) calls with client.messages.create(...) or client.messages.stream(...).
2. Use the helper utilities to create intermediate function_call / function_call_result messages when you execute tools locally.
3. Profit! No other structural changes are required.

---

Testing & Development

# install dev dependencies
pip install -r requirements.txt

# run unit tests (add more as the SDK grows)
pytest

# run lint/format tools if configured (e.g. ruff, black)

The repository ships with a .venv/ suggestion for isolation; feel free to use your own environment manager.

---

Contributing

Pull requests are welcome! Please:

1. Fork & clone the repository.
2. Create a topic branch (git checkout -b feat/awesome-improvement).
3. Add tests for new functionality when possible.
4. Open a PR describing the change and cookbook scenarios it improves.

For major changes, open an issue first to discuss what you’d like to see.

---

License

MIT License. See LICENSE.

---

Integrations

The SDK wraps the /v1/integrations endpoints so you can enumerate, connect, and execute third-party apps with just a few lines of code.

client = Incredible()

# List available integrations
for integration in client.integrations.list():
    print(integration["id"], integration["name"])

# Fetch details for a specific integration
perplexity = client.integrations.retrieve("perplexity")
print(perplexity["features"])

# Connect a user (API key example)
client.integrations.connect(
    "perplexity",
    user_id="user_123",
    api_key="perplexity-secret",
)

# Execute a feature from a connected integration
result = client.integrations.execute(
    "perplexity",
    user_id="user_123",
    feature_name="PERPLEXITY_SEARCH",
    inputs={"query": "Latest AI news"},
)
print(result)

Supports both API key and OAuth flows via the same connect call—pass api_key or callback_url depending on the integration’s auth_method.

---

Reference & Helpers

The SDK ships with helper utilities so you can mirror the cookbook patterns without manual parsing. Key modules:

• helpers.build_tool_execution_plan(response.raw) → parses the model’s requested tool calls.
• helpers.execute_plan(plan, registry) → runs your Python callables in order, returns results.
• helpers.build_follow_up_messages(messages, plan, results) → generates the follow-up function_call and function_call_result messages you can feed back to the model.

Example:

from incredible_python import Incredible, helpers

client = Incredible()
response = client.messages.create(
    model="small-1",
    max_tokens=256,
    messages=[{"role": "user", "content": "Add 7 and 35, then look up [e-mail]"}],
    functions=get_tool_schemas(),
)

plan = helpers.build_tool_execution_plan(response.raw)
if plan and not plan.is_empty():
    results = helpers.execute_plan(plan, registry=get_tool_registry())
    follow_up = helpers.build_follow_up_messages(response.raw["result"]["messages"], plan, results)
    final = client.messages.create(
        model="small-1",
        max_tokens=256,
        messages=follow_up,
        functions=get_tool_schemas(),
    )

See docs/python_sdk_helpers.md for full argument tables and advanced patterns such as wrapping the entire flow inside a reusable run_agent_query() helper.