Claudette’s source

This is the ‘literate’ source code for Claudette. You can view the fully rendered version of the notebook here, or you can clone the git repo and run the interactive notebook in Jupyter. The notebook is converted the Python module claudette/core.py using nbdev. The goal of this source code is to both create the Python module, and also to teach the reader how it is created, without assuming much existing knowledge about Claude’s API.

Most of the time you’ll see that we write some source code first, and then a description or discussion of it afterwards.

Setup

import os
# os.environ['ANTHROPIC_LOG'] = 'debug'

To print every HTTP request and response in full, uncomment the above line. This functionality is provided by Anthropic’s SDK.

Tip

If you’re reading the rendered version of this notebook, you’ll see an “Exported source” collapsible widget below. If you’re reading the source notebook directly, you’ll see #| exports at the top of the cell. These show that this piece of code will be exported into the python module that this notebook creates. No other code will be included – any other code in this notebook is just for demonstration, documentation, and testing.

You can toggle expanding/collapsing the source code of all exported sections by using the </> Code menu in the top right of the rendered notebook page.

Exported source 
model_types = {
    # Anthropic
    'claude-3-opus-20240229': 'opus',
    'claude-3-5-sonnet-20241022': 'sonnet',
    'claude-3-haiku-20240307': 'haiku-3',
    'claude-3-5-haiku-20241022': 'haiku-3-5',
    # AWS
    'anthropic.claude-3-opus-20240229-v1:0': 'opus',
    'anthropic.claude-3-5-sonnet-20241022-v2:0': 'sonnet',
    'anthropic.claude-3-sonnet-20240229-v1:0': 'sonnet',
    'anthropic.claude-3-haiku-20240307-v1:0': 'haiku',
    # Google
    'claude-3-opus@20240229': 'opus',
    'claude-3-5-sonnet-v2@20241022': 'sonnet',
    'claude-3-sonnet@20240229': 'sonnet',
    'claude-3-haiku@20240307': 'haiku',
}

all_models = list(model_types)
Exported source 
text_only_models = ('claude-3-5-haiku-20241022',)

These are the current versions and prices of Anthropic’s models at the time of writing.

model = models[1]; model
'claude-3-5-sonnet-20241022'

For examples, we’ll use Sonnet 3.5, since it’s awesome.

Antropic SDK

cli = Anthropic()

This is what Anthropic’s SDK provides for interacting with Python. To use it, pass it a list of messages, with content and a role. The roles should alternate between user and assistant.

Tip

After the code below you’ll see an indented section with an orange vertical line on the left. This is used to show the result of running the code above. Because the code is running in a Jupyter Notebook, we don’t have to use print to display results, we can just type the expression directly, as we do with r here.

m = {'role': 'user', 'content': "I'm Jeremy"}
r = cli.messages.create(messages=[m], model=model, max_tokens=100)
r

Hi Jeremy! Nice to meet you. I’m Claude. How can I help you today?

  • id: msg_01HJ7w89W8bFTMeWhWCe1ZGZ
  • content: [{'text': "Hi Jeremy! Nice to meet you. I'm Claude. How can I help you today?", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 10, 'output_tokens': 22}

Formatting output

That output is pretty long and hard to read, so let’s clean it up. We’ll start by pulling out the Content part of the message. To do that, we’re going to write our first function which will be included to the claudette/core.py module.

Tip

This is the first exported public function or class we’re creating (the previous export was of a variable). In the rendered version of the notebook for these you’ll see 4 things, in this order (unless the symbol starts with a single _, which indicates it’s private):

  • The signature (with the symbol name as a heading, with a horizontal rule above)
  • A table of paramater docs (if provided)
  • The doc string (in italics).
  • The source code (in a collapsible “Exported source” block)

After that, we generally provide a bit more detail on what we’ve created, and why, along with a sample usage.

source

find_block

 find_block (r:collections.abc.Mapping, blk_type:type=<class
             'anthropic.types.text_block.TextBlock'>)

Find the first block of type blk_type in r.content.

Type Default Details
r Mapping The message to look in
blk_type type TextBlock The type of block to find
Exported source 
def find_block(r:abc.Mapping, # The message to look in
               blk_type:type=TextBlock  # The type of block to find
              ):
    "Find the first block of type `blk_type` in `r.content`."
    return first(o for o in r.content if isinstance(o,blk_type))

This makes it easier to grab the needed parts of Claude’s responses, which can include multiple pieces of content. By default, we look for the first text block. That will generally have the content we want to display.

find_block(r)
TextBlock(text="Hi Jeremy! Nice to meet you. I'm Claude. How can I help you today?", type='text')

source

contents

 contents (r)

Helper to get the contents from Claude response r.

Exported source 
def contents(r):
    "Helper to get the contents from Claude response `r`."
    blk = find_block(r)
    if not blk and r.content: blk = r.content[0]
    return blk.text.strip() if hasattr(blk,'text') else str(blk)

For display purposes, we often just want to show the text itself.

contents(r)
"Hi Jeremy! Nice to meet you. I'm Claude. How can I help you today?"
Exported source 
@patch
def _repr_markdown_(self:(Message)):
    det = '\n- '.join(f'{k}: `{v}`' for k,v in self.model_dump().items())
    cts = re.sub(r'\$', '&#36;', contents(self))  # escape `$` for jupyter latex
    return f"""{cts}

<details>

- {det}

</details>"""

Jupyter looks for a _repr_markdown_ method in displayed objects; we add this in order to display just the content text, and collapse full details into a hideable section. Note that patch is from fastcore, and is used to add (or replace) functionality in an existing class. We pass the class(es) that we want to patch as type annotations to self. In this case, _repr_markdown_ is being added to Anthropic’s Message class, so when we display the message now we just see the contents, and the details are hidden away in a collapsible details block.

r

Hi Jeremy! Nice to meet you. I’m Claude. How can I help you today?

  • id: msg_01HJ7w89W8bFTMeWhWCe1ZGZ
  • content: [{'text': "Hi Jeremy! Nice to meet you. I'm Claude. How can I help you today?", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 10, 'output_tokens': 22}

One key part of the response is the usage key, which tells us how many tokens we used by returning a Usage object.

We’ll add some helpers to make things a bit cleaner for creating and formatting these objects.

r.usage
In: 10; Out: 22; Cache create: 0; Cache read: 0; Total: 32

source

usage

 usage (inp=0, out=0, cache_create=0, cache_read=0)

Slightly more concise version of Usage.

Type Default Details
inp int 0 input tokens
out int 0 Output tokens
cache_create int 0 Cache creation tokens
cache_read int 0 Cache read tokens
Exported source 
def usage(inp=0, # input tokens
          out=0,  # Output tokens
          cache_create=0, # Cache creation tokens
          cache_read=0 # Cache read tokens
         ):
    "Slightly more concise version of `Usage`."
    return Usage(input_tokens=inp, output_tokens=out, cache_creation_input_tokens=cache_create, cache_read_input_tokens=cache_read)

The constructor provided by Anthropic is rather verbose, so we clean it up a bit, using a lowercase version of the name.

usage(5)
In: 5; Out: 0; Cache create: 0; Cache read: 0; Total: 5

source

Usage.total

 Usage.total ()
Exported source 
@patch(as_prop=True)
def total(self:Usage): return self.input_tokens+self.output_tokens+getattr(self, "cache_creation_input_tokens",0)+getattr(self, "cache_read_input_tokens",0)

Adding a total property to Usage makes it easier to see how many tokens we’ve used up altogether.

usage(5,1).total
6

source

Usage.__repr__

 Usage.__repr__ ()

Return repr(self).

Exported source 
@patch
def __repr__(self:Usage): return f'In: {self.input_tokens}; Out: {self.output_tokens}; Cache create: {getattr(self, "cache_creation_input_tokens",0)}; Cache read: {getattr(self, "cache_read_input_tokens",0)}; Total: {self.total}'

In python, patching __repr__ lets us change how an object is displayed. (More generally, methods starting and ending in __ in Python are called dunder methods, and have some magic behavior – such as, in this case, changing how an object is displayed.)

usage(5)
In: 5; Out: 0; Cache create: 0; Cache read: 0; Total: 5

source

Usage.__add__

 Usage.__add__ (b)

Add together each of input_tokens and output_tokens

Exported source 
@patch
def __add__(self:Usage, b):
    "Add together each of `input_tokens` and `output_tokens`"
    return usage(self.input_tokens+b.input_tokens, self.output_tokens+b.output_tokens, getattr(self,'cache_creation_input_tokens',0)+getattr(b,'cache_creation_input_tokens',0), getattr(self,'cache_read_input_tokens',0)+getattr(b,'cache_read_input_tokens',0))

And, patching __add__ lets + work on a Usage object.

r.usage+r.usage
In: 20; Out: 44; Cache create: 0; Cache read: 0; Total: 64

Creating messages

Creating correctly formatted dicts from scratch every time isn’t very handy, so we’ll import a couple of helper functions from the msglm library.

Let’s use mk_msg to recreate our msg {'role': 'user', 'content': "I'm Jeremy"} from earlier.

prompt = "I'm Jeremy"
m = mk_msg(prompt)
r = cli.messages.create(messages=[m], model=model, max_tokens=100)
r

Hi Jeremy! Nice to meet you. I’m Claude, an AI assistant. How can I help you today?

  • id: msg_011ufPxz2zWw7QH9Wgb57XR8
  • content: [{'text': "Hi Jeremy! Nice to meet you. I'm Claude, an AI assistant. How can I help you today?", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 10, 'output_tokens': 26}

We can pass more than just text messages to Claude. As we’ll see later we can also pass images, SDK objects, etc. To handle these different data types we need to pass the type along with our content to Claude.

Here’s an example of a multimodal message containing text and images.

{
    'role': 'user', 
    'content': [
        {'type':'text', 'text':'What is in the image?'},
        {
            'type':'image', 
            'source': {
                'type':'base64', 'media_type':'media_type', 'data': 'data'
            }
        }
    ]
}

mk_msg infers the type automatically and creates the appropriate data structure.

LLMs, don’t actually have state, but instead dialogs are created by passing back all previous prompts and responses every time. With Claude, they always alternate user and assistant. We’ll use mk_msgs from msglm to make it easier to build up these dialog lists.

msgs = mk_msgs([prompt, r, "I forgot my name. Can you remind me please?"]) 
msgs
[{'role': 'user', 'content': "I'm Jeremy"},
 {'role': 'assistant',
  'content': [TextBlock(text="Hi Jeremy! Nice to meet you. I'm Claude, an AI assistant. How can I help you today?", type='text')]},
 {'role': 'user', 'content': 'I forgot my name. Can you remind me please?'}]
cli.messages.create(messages=msgs, model=model, max_tokens=200)

You just told me your name is Jeremy.

  • id: msg_01EPqVxLZ84BV8JqsahuBncU
  • content: [{'text': 'You just told me your name is Jeremy.', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 50, 'output_tokens': 12}

Client

source

Client

 Client (model, cli=None, log=False)

Basic Anthropic messages client.

Exported source 
class Client:
    def __init__(self, model, cli=None, log=False):
        "Basic Anthropic messages client."
        self.model,self.use = model,usage()
        self.text_only = model in text_only_models
        self.log = [] if log else None
        self.c = (cli or Anthropic(default_headers={'anthropic-beta': 'prompt-caching-2024-07-31'}))

We’ll create a simple Client for Anthropic which tracks usage stores the model to use. We don’t add any methods right away – instead we’ll use patch for that so we can add and document them incrementally.

c = Client(model)
c.use
In: 0; Out: 0; Cache create: 0; Cache read: 0; Total: 0
Exported source 
@patch
def _r(self:Client, r:Message, prefill=''):
    "Store the result of the message and accrue total usage."
    if prefill:
        blk = find_block(r)
        blk.text = prefill + (blk.text or '')
    self.result = r
    self.use += r.usage
    self.stop_reason = r.stop_reason
    self.stop_sequence = r.stop_sequence
    return r

We use a _ prefix on private methods, but we document them here in the interests of literate source code.

_r will be used each time we get a new result, to track usage and also to keep the result available for later.

c._r(r)
c.use
In: 10; Out: 26; Cache create: 0; Cache read: 0; Total: 36

Whereas OpenAI’s models use a stream parameter for streaming, Anthropic’s use a separate method. We implement Anthropic’s approach in a private method, and then use a stream parameter in __call__ for consistency:

Exported source 
@patch
def _log(self:Client, final, prefill, msgs, maxtok=None, sp=None, temp=None, stream=None, stop=None, **kwargs):
    self._r(final, prefill)
    if self.log is not None: self.log.append({
        "msgs": msgs, "prefill": prefill, **kwargs,
        "msgs": msgs, "prefill": prefill, "maxtok": maxtok, "sp": sp, "temp": temp, "stream": stream, "stop": stop, **kwargs,
        "result": self.result, "use": self.use, "stop_reason": self.stop_reason, "stop_sequence": self.stop_sequence
    })
    return self.result
Exported source 
@patch
def _stream(self:Client, msgs:list, prefill='', **kwargs):
    with self.c.messages.stream(model=self.model, messages=mk_msgs(msgs), **kwargs) as s:
        if prefill: yield(prefill)
        yield from s.text_stream
        self._log(s.get_final_message(), prefill, msgs, **kwargs)

Claude supports adding an extra assistant message at the end, which contains the prefill – i.e. the text we want Claude to assume the response starts with. However Claude doesn’t actually repeat that in the response, so for convenience we add it.

Exported source 
@patch
def _precall(self:Client, msgs, prefill, stop, kwargs):
    pref = [prefill.strip()] if prefill else []
    if not isinstance(msgs,list): msgs = [msgs]
    if stop is not None:
        if not isinstance(stop, (list)): stop = [stop]
        kwargs["stop_sequences"] = stop
    msgs = mk_msgs(msgs+pref)
    return msgs
@patch
@delegates(messages.Messages.create)
def __call__(self:Client,
             msgs:list, # List of messages in the dialog
             sp='', # The system prompt
             temp=0, # Temperature
             maxtok=4096, # Maximum tokens
             prefill='', # Optional prefill to pass to Claude as start of its response
             stream:bool=False, # Stream response?
             stop=None, # Stop sequence
             **kwargs):
    "Make a call to Claude."
    msgs = self._precall(msgs, prefill, stop, kwargs)
    if stream: return self._stream(msgs, prefill=prefill, max_tokens=maxtok, system=sp, temperature=temp, **kwargs)
    res = self.c.messages.create(
        model=self.model, messages=msgs, max_tokens=maxtok, system=sp, temperature=temp, **kwargs)
    return self._log(res, prefill, msgs, maxtok, sp, temp, stream=stream, **kwargs)

Defining __call__ let’s us use an object like a function (i.e it’s callable). We use it as a small wrapper over messages.create. However we’re not exporting this version just yet – we have some additions we’ll make in a moment…

c = Client(model, log=True)
c.use
In: 0; Out: 0; Cache create: 0; Cache read: 0; Total: 0
c('Hi')

Hello! How can I help you today?

  • id: msg_01HV3iasP1oA5qefPPhKcduZ
  • content: [{'text': 'Hello! How can I help you today?', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 8, 'output_tokens': 12, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}
c.use
In: 8; Out: 12; Cache create: 0; Cache read: 0; Total: 20

Let’s try out prefill:

q = "Concisely, what is the meaning of life?"
pref = 'According to Douglas Adams,'
c(q, prefill=pref)

According to Douglas Adams, it’s 42. More seriously, there’s no universal answer - it’s deeply personal. Common perspectives include: finding happiness, making meaningful connections, pursuing purpose through work/creativity, helping others, or simply experiencing and appreciating existence.

  • id: msg_014swwsm5y9HtezY4VjShyXu
  • content: [{'text': "According to Douglas Adams, it's 42. More seriously, there's no universal answer - it's deeply personal. Common perspectives include: finding happiness, making meaningful connections, pursuing purpose through work/creativity, helping others, or simply experiencing and appreciating existence.", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 24, 'output_tokens': 52, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

We can pass stream=True to stream the response back incrementally:

for o in c('Hi', stream=True): print(o, end='')
Hello! How can I help you today?
c.use
In: 40; Out: 76; Cache create: 0; Cache read: 0; Total: 116
for o in c(q, prefill=pref, stream=True): print(o, end='')
According to Douglas Adams,  it's 42. More seriously, there's no universal answer - it's deeply personal. Common perspectives include: finding happiness, creating meaning through relationships and achievements, pursuing knowledge, helping others, or fulfilling one's potential.
c.use
In: 64; Out: 126; Cache create: 0; Cache read: 0; Total: 190

Pass a stop seauence if you want claude to stop generating text when it encounters it.

c("Count from 1 to 10", stop="5")

1 2 3 4

  • id: msg_01EkMdim6ZVZHNyqWYNVaQac
  • content: [{'text': '1\n2\n3\n4\n', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: stop_sequence
  • stop_sequence: 5
  • type: message
  • usage: {'input_tokens': 15, 'output_tokens': 10, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

This also works with streaming, and you can pass more than one stop sequence:

for o in c("Count from 1 to 10", stop=["2", "yellow"], stream=True): print(o, end='')
print(c.stop_reason, c.stop_sequence)
1
stop_sequence 2

You can check the logs:

c.log[-1]
{'msgs': [{'role': 'user', 'content': 'Count from 1 to 10'}],
 'prefill': '',
 'max_tokens': 4096,
 'system': '',
 'temperature': 0,
 'stop_sequences': ['2', 'yellow'],
 'maxtok': None,
 'sp': None,
 'temp': None,
 'stream': None,
 'stop': None,
 'result': Message(id='msg_01StaLq6MXV7W1nfQM3Tcr4o', content=[TextBlock(text='1\n', type='text')], model='claude-3-5-sonnet-20241022', role='assistant', stop_reason='stop_sequence', stop_sequence='2', type='message', usage=In: 15; Out: 4; Cache create: 0; Cache read: 0; Total: 19),
 'use': In: 94; Out: 140; Cache create: 0; Cache read: 0; Total: 234,
 'stop_reason': 'stop_sequence',
 'stop_sequence': '2'}

Tool use

Let’s now add tool use (aka function calling).

source

mk_tool_choice

 mk_tool_choice (choose:Union[str,bool,NoneType])

Create a tool_choice dict that’s ‘auto’ if choose is None, ‘any’ if it is True, or ‘tool’ otherwise

Exported source 
def mk_tool_choice(choose:Union[str,bool,None])->dict:
    "Create a `tool_choice` dict that's 'auto' if `choose` is `None`, 'any' if it is True, or 'tool' otherwise"
    return {"type": "tool", "name": choose} if isinstance(choose,str) else {'type':'any'} if choose else {'type':'auto'}
print(mk_tool_choice('sums'))
print(mk_tool_choice(True))
print(mk_tool_choice(None))
{'type': 'tool', 'name': 'sums'}
{'type': 'any'}
{'type': 'auto'}

Claude can be forced to use a particular tool, or select from a specific list of tools, or decide for itself when to use a tool. If you want to force a tool (or force choosing from a list), include a tool_choice param with a dict from mk_tool_choice.

For testing, we need a function that Claude can call; we’ll write a simple function that adds numbers together, and will tell us when it’s being called:

def sums(
    a:int,  # First thing to sum
    b:int=1 # Second thing to sum
) -> int: # The sum of the inputs
    "Adds a + b."
    print(f"Finding the sum of {a} and {b}")
    return a + b
a,b = 604542,6458932
pr = f"What is {a}+{b}?"
sp = "You are a summing expert."

Claudette can autogenerate a schema thanks to the toolslm library. We’ll force the use of the tool using the function we created earlier.

tools=[get_schema(sums)]
choice = mk_tool_choice('sums')

We’ll start a dialog with Claude now. We’ll store the messages of our dialog in msgs. The first message will be our prompt pr, and we’ll pass our tools schema.

msgs = mk_msgs(pr)
r = c(msgs, sp=sp, tools=tools, tool_choice=choice)
r

ToolUseBlock(id=‘toolu_01SU3EQ2ypNHev8Xos38C3Xe’, input={‘a’: 604542, ‘b’: 6458932}, name=‘sums’, type=‘tool_use’)

  • id: msg_01EphS1Wj7X9714vamCFYqak
  • content: [{'id': 'toolu_01SU3EQ2ypNHev8Xos38C3Xe', 'input': {'a': 604542, 'b': 6458932}, 'name': 'sums', 'type': 'tool_use'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: tool_use
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 442, 'output_tokens': 53, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

When Claude decides that it should use a tool, it passes back a ToolUseBlock with the name of the tool to call, and the params to use.

We don’t want to allow it to call just any possible function (that would be a security disaster!) so we create a namespace – that is, a dictionary of allowable function names to call.

ns = mk_ns(sums)
ns
{'sums': <function __main__.sums(a: int, b: int = 1) -> int>}

source

mk_funcres

 mk_funcres (tuid, res)

Given tool use id and the tool result, create a tool_result response.

Exported source 
def mk_funcres(tuid, res):
    "Given tool use id and the tool result, create a tool_result response."
    return dict(type="tool_result", tool_use_id=tuid, content=str(res))

We can now use the function requested by Claude. We look it up in ns, and pass in the provided parameters.

fc = find_block(r, ToolUseBlock)
res = mk_funcres(fc.id, call_func(fc.name, fc.input, ns=ns))
res
Finding the sum of 604542 and 6458932
{'type': 'tool_result',
 'tool_use_id': 'toolu_01SU3EQ2ypNHev8Xos38C3Xe',
 'content': '7063474'}

source

mk_toolres

 mk_toolres (r:collections.abc.Mapping,
             ns:Optional[collections.abc.Mapping]=None, obj:Optional=None)

Create a tool_result message from response r.

Type Default Details
r Mapping Tool use request response from Claude
ns Optional None Namespace to search for tools
obj Optional None Class to search for tools
Exported source 
def mk_toolres(
    r:abc.Mapping, # Tool use request response from Claude
    ns:Optional[abc.Mapping]=None, # Namespace to search for tools
    obj:Optional=None # Class to search for tools
    ):
    "Create a `tool_result` message from response `r`."
    cts = getattr(r, 'content', [])
    res = [mk_msg(r)]
    if ns is None: ns=globals()
    if obj is not None: ns = mk_ns(obj)
    tcs = [mk_funcres(o.id, call_func(o.name, o.input, ns)) for o in cts if isinstance(o,ToolUseBlock)]
    if tcs: res.append(mk_msg(tcs))
    return res

In order to tell Claude the result of the tool call, we pass back the tool use assistant request and the tool_result response.

tr = mk_toolres(r, ns=ns)
tr
Finding the sum of 604542 and 6458932
[{'role': 'assistant',
  'content': [ToolUseBlock(id='toolu_01SU3EQ2ypNHev8Xos38C3Xe', input={'a': 604542, 'b': 6458932}, name='sums', type='tool_use')]},
 {'role': 'user',
  'content': [{'type': 'tool_result',
    'tool_use_id': 'toolu_01SU3EQ2ypNHev8Xos38C3Xe',
    'content': '7063474'}]}]

We add this to our dialog, and now Claude has all the information it needs to answer our question.

msgs += tr
contents(c(msgs, sp=sp, tools=tools))
'The sum of 604542 and 6458932 is 7063474.'
msgs
[{'role': 'user', 'content': 'What is 604542+6458932?'},
 {'role': 'assistant',
  'content': [ToolUseBlock(id='toolu_01SU3EQ2ypNHev8Xos38C3Xe', input={'a': 604542, 'b': 6458932}, name='sums', type='tool_use')]},
 {'role': 'user',
  'content': [{'type': 'tool_result',
    'tool_use_id': 'toolu_01SU3EQ2ypNHev8Xos38C3Xe',
    'content': '7063474'}]}]

This works with methods as well – in this case, use the object itself for ns:

class Dummy:
    def sums(
        self,
        a:int,  # First thing to sum
        b:int=1 # Second thing to sum
    ) -> int: # The sum of the inputs
        "Adds a + b."
        print(f"Finding the sum of {a} and {b}")
        return a + b
tools = [get_schema(Dummy.sums)]
o = Dummy()
r = c(pr, sp=sp, tools=tools, tool_choice=choice)
tr = mk_toolres(r, obj=o)
msgs += tr
contents(c(msgs, sp=sp, tools=tools))
Finding the sum of 604542 and 6458932
'The sum of 604542 and 6458932 is 7063474.'

source

get_types

 get_types (msgs)
get_types(msgs)
['text', 'tool_use', 'tool_result', 'tool_use', 'tool_result']

source

Client.__call__

 Client.__call__ (msgs:list, sp='', temp=0, maxtok=4096, prefill='',
                  stream:bool=False, stop=None, tools:Optional[list]=None,
                  tool_choice:Optional[dict]=None,
                  metadata:MetadataParam|NotGiven=NOT_GIVEN,
                  stop_sequences:List[str]|NotGiven=NOT_GIVEN, system:Unio
                  n[str,Iterable[TextBlockParam]]|NotGiven=NOT_GIVEN,
                  temperature:float|NotGiven=NOT_GIVEN,
                  top_k:int|NotGiven=NOT_GIVEN,
                  top_p:float|NotGiven=NOT_GIVEN,
                  extra_headers:Headers|None=None,
                  extra_query:Query|None=None, extra_body:Body|None=None,
                  timeout:float|httpx.Timeout|None|NotGiven=NOT_GIVEN)

Make a call to Claude.

Type Default Details
msgs list List of messages in the dialog
sp str The system prompt
temp int 0 Temperature
maxtok int 4096 Maximum tokens
prefill str Optional prefill to pass to Claude as start of its response
stream bool False Stream response?
stop NoneType None Stop sequence
tools Optional None List of tools to make available to Claude
tool_choice Optional None Optionally force use of some tool
metadata MetadataParam | NotGiven NOT_GIVEN
stop_sequences List[str] | NotGiven NOT_GIVEN
system Union[str, Iterable[TextBlockParam]] | NotGiven NOT_GIVEN
temperature float | NotGiven NOT_GIVEN
top_k int | NotGiven NOT_GIVEN
top_p float | NotGiven NOT_GIVEN
extra_headers Headers | None None
extra_query Query | None None
extra_body Body | None None
timeout float | httpx.Timeout | None | NotGiven NOT_GIVEN
Exported source 
@patch
@delegates(messages.Messages.create)
def __call__(self:Client,
             msgs:list, # List of messages in the dialog
             sp='', # The system prompt
             temp=0, # Temperature
             maxtok=4096, # Maximum tokens
             prefill='', # Optional prefill to pass to Claude as start of its response
             stream:bool=False, # Stream response?
             stop=None, # Stop sequence
             tools:Optional[list]=None, # List of tools to make available to Claude
             tool_choice:Optional[dict]=None, # Optionally force use of some tool
             **kwargs):
    "Make a call to Claude."
    if tools: kwargs['tools'] = [get_schema(o) for o in listify(tools)]
    if tool_choice: kwargs['tool_choice'] = mk_tool_choice(tool_choice)
    msgs = self._precall(msgs, prefill, stop, kwargs)
    if any(t == 'image' for t in get_types(msgs)): assert not self.text_only, f"Images are not supported by the current model type: {self.model}"
    if stream: return self._stream(msgs, prefill=prefill, max_tokens=maxtok, system=sp, temperature=temp, **kwargs)
    res = self.c.messages.create(model=self.model, messages=msgs, max_tokens=maxtok, system=sp, temperature=temp, **kwargs)
    return self._log(res, prefill, msgs, maxtok, sp, temp, stream=stream, stop=stop, **kwargs)
r = c(pr, sp=sp, tools=sums, tool_choice=sums)
r

ToolUseBlock(id=‘toolu_018aevns2KSCEn4LCL7zNX2n’, input={‘a’: 604542, ‘b’: 6458932}, name=‘sums’, type=‘tool_use’)

  • id: msg_01JqjXtvM3j9YdDG7gN9vvym
  • content: [{'id': 'toolu_018aevns2KSCEn4LCL7zNX2n', 'input': {'a': 604542, 'b': 6458932}, 'name': 'sums', 'type': 'tool_use'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: tool_use
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 438, 'output_tokens': 57, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}
tr = mk_toolres(r, ns=ns)
Finding the sum of 604542 and 6458932

source

Client.structured

 Client.structured (msgs:list, tools:Optional[list]=None,
                    obj:Optional=None,
                    ns:Optional[collections.abc.Mapping]=None, sp='',
                    temp=0, maxtok=4096, prefill='', stream:bool=False,
                    stop=None, tool_choice:Optional[dict]=None,
                    metadata:MetadataParam|NotGiven=NOT_GIVEN,
                    stop_sequences:List[str]|NotGiven=NOT_GIVEN, system:Un
                    ion[str,Iterable[TextBlockParam]]|NotGiven=NOT_GIVEN,
                    temperature:float|NotGiven=NOT_GIVEN,
                    top_k:int|NotGiven=NOT_GIVEN,
                    top_p:float|NotGiven=NOT_GIVEN,
                    extra_headers:Headers|None=None,
                    extra_query:Query|None=None,
                    extra_body:Body|None=None,
                    timeout:float|httpx.Timeout|None|NotGiven=NOT_GIVEN)

Return the value of all tool calls (generally used for structured outputs)

Type Default Details
msgs list List of messages in the dialog
tools Optional None List of tools to make available to Claude
obj Optional None Class to search for tools
ns Optional None Namespace to search for tools
sp str The system prompt
temp int 0 Temperature
maxtok int 4096 Maximum tokens
prefill str Optional prefill to pass to Claude as start of its response
stream bool False Stream response?
stop NoneType None Stop sequence
tool_choice Optional None Optionally force use of some tool
metadata MetadataParam | NotGiven NOT_GIVEN
stop_sequences List[str] | NotGiven NOT_GIVEN
system Union[str, Iterable[TextBlockParam]] | NotGiven NOT_GIVEN
temperature float | NotGiven NOT_GIVEN
top_k int | NotGiven NOT_GIVEN
top_p float | NotGiven NOT_GIVEN
extra_headers Headers | None None
extra_query Query | None None
extra_body Body | None None
timeout float | httpx.Timeout | None | NotGiven NOT_GIVEN
Exported source 
@patch
@delegates(Client.__call__)
def structured(self:Client,
               msgs:list, # List of messages in the dialog
               tools:Optional[list]=None, # List of tools to make available to Claude
               obj:Optional=None, # Class to search for tools
               ns:Optional[abc.Mapping]=None, # Namespace to search for tools
               **kwargs):
    "Return the value of all tool calls (generally used for structured outputs)"
    tools = listify(tools)
    res = self(msgs, tools=tools, tool_choice=tools, **kwargs)
    if ns is None: ns=mk_ns(*tools)
    if obj is not None: ns = mk_ns(obj)
    cts = getattr(res, 'content', [])
    tcs = [call_func(o.name, o.input, ns=ns) for o in cts if isinstance(o,ToolUseBlock)]
    return tcs

Anthropic’s API does not support response formats directly, so instead we provide a structured method to use tool calling to achieve the same result. The result of the tool is not passed back to Claude in this case, but instead is returned directly to the user.

c.structured(pr, tools=[sums])
Finding the sum of 604542 and 6458932
[7063474]

Chat

Rather than manually adding the responses to a dialog, we’ll create a simple Chat class to do that for us, each time we make a request. We’ll also store the system prompt and tools here, to avoid passing them every time.

source

Chat

 Chat (model:Optional[str]=None, cli:Optional[__main__.Client]=None,
       sp='', tools:Optional[list]=None, temp=0,
       cont_pr:Optional[str]=None)

Anthropic chat client.

Type Default Details
model Optional None Model to use (leave empty if passing cli)
cli Optional None Client to use (leave empty if passing model)
sp str Optional system prompt
tools Optional None List of tools to make available to Claude
temp int 0 Temperature
cont_pr Optional None User prompt to continue an assistant response: assistant,[user:“…”],assistant
Exported source 
class Chat:
    def __init__(self,
                 model:Optional[str]=None, # Model to use (leave empty if passing `cli`)
                 cli:Optional[Client]=None, # Client to use (leave empty if passing `model`)
                 sp='', # Optional system prompt
                 tools:Optional[list]=None, # List of tools to make available to Claude
                 temp=0, # Temperature
                 cont_pr:Optional[str]=None): # User prompt to continue an assistant response: assistant,[user:"..."],assistant
        "Anthropic chat client."
        assert model or cli
        assert cont_pr != "", "cont_pr may not be an empty string"
        self.c = (cli or Client(model))
        self.h,self.sp,self.tools,self.cont_pr,self.temp = [],sp,tools,cont_pr,temp

    @property
    def use(self): return self.c.use

The class stores the Client that will provide the responses in c, and a history of messages in h.

sp = "Never mention what tools you use."
chat = Chat(model, sp=sp)
chat.c.use, chat.h
(In: 0; Out: 0; Cache create: 0; Cache read: 0; Total: 0, [])

We’ve shown the token usage but we really care about is pricing. Let’s extract the latest pricing from Anthropic into a pricing dict.

We’ll patch Usage to enable it compute the cost given pricing.

source

Usage.cost

 Usage.cost (costs:tuple)
Exported source 
@patch
def cost(self:Usage, costs:tuple) -> float:
    cache_w, cache_r = getattr(self, "cache_creation_input_tokens",0), getattr(self, "cache_read_input_tokens",0)
    return sum([self.input_tokens * costs[0] +  self.output_tokens * costs[1] +  cache_w * costs[2] + cache_r * costs[3]]) / 1e6
chat.c.use.cost(pricing[model_types[chat.c.model]])
0.0

This is clunky. Let’s add cost as a property for the Chat class. It will pass in the appropriate prices for the current model to the usage cost calculator.

source

Chat.cost

 Chat.cost ()
Exported source 
@patch(as_prop=True)
def cost(self: Chat) -> float: return self.c.use.cost(pricing[model_types[self.c.model]])
chat.cost
0.0

source

Chat.__call__

 Chat.__call__ (pr=None, temp=None, maxtok=4096, stream=False, prefill='',
                tool_choice:Optional[dict]=None, **kw)

Call self as a function.

Type Default Details
pr NoneType None Prompt / message
temp NoneType None Temperature
maxtok int 4096 Maximum tokens
stream bool False Stream response?
prefill str Optional prefill to pass to Claude as start of its response
tool_choice Optional None Optionally force use of some tool
kw
Exported source 
@patch
def _stream(self:Chat, res):
    yield from res
    self.h += mk_toolres(self.c.result, ns=self.tools, obj=self)
Exported source 
@patch
def _post_pr(self:Chat, pr, prev_role):
    if pr is None and prev_role == 'assistant':
        if self.cont_pr is None:
            raise ValueError("Prompt must be given after assistant completion, or use `self.cont_pr`.")
        pr = self.cont_pr # No user prompt, keep the chain
    if pr: self.h.append(mk_msg(pr))
Exported source 
@patch
def _append_pr(self:Chat,
               pr=None,  # Prompt / message
              ):
    prev_role = nested_idx(self.h, -1, 'role') if self.h else 'assistant' # First message should be 'user'
    if pr and prev_role == 'user': self() # already user request pending
    self._post_pr(pr, prev_role)
Exported source 
@patch
def __call__(self:Chat,
             pr=None,  # Prompt / message
             temp=None, # Temperature
             maxtok=4096, # Maximum tokens
             stream=False, # Stream response?
             prefill='', # Optional prefill to pass to Claude as start of its response
             tool_choice:Optional[dict]=None, # Optionally force use of some tool
             **kw):
    if temp is None: temp=self.temp
    self._append_pr(pr)
    res = self.c(self.h, stream=stream, prefill=prefill, sp=self.sp, temp=temp, maxtok=maxtok,
                 tools=self.tools, tool_choice=tool_choice,**kw)
    if stream: return self._stream(res)
    self.h += mk_toolres(self.c.result, ns=self.tools)
    return res

The __call__ method just passes the request along to the Client, but rather than just passing in this one prompt, it appends it to the history and passes it all along. As a result, we now have state!

chat = Chat(model, sp=sp)
chat("I'm Jeremy")
chat("What's my name?")

Your name is Jeremy.

  • id: msg_01Mm7Fxka3mcEzCWjq8XnHnP
  • content: [{'text': 'Your name is Jeremy.', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 41, 'output_tokens': 8, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}
chat.use, chat.cost
(In: 58; Out: 24; Cache create: 0; Cache read: 0; Total: 82, 0.000534)

Let’s try out prefill too:

q = "Concisely, what is the meaning of life?"
pref = 'According to Douglas Adams,'
chat(q, prefill=pref)

According to Douglas Adams, 42. But in reality, it’s to find personal meaning through experiences, relationships, and pursuing what brings you fulfillment.

  • id: msg_012rrj71bX5ePoqpQxDCHkzH
  • content: [{'text': "According to Douglas Adams, 42. But in reality, it's to find personal meaning through experiences, relationships, and pursuing what brings you fulfillment.", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 69, 'output_tokens': 30, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

By default messages must be in user, assistant, user format. If this isn’t followed (aka calling chat() without a user message) it will error out:

try: chat()
except ValueError as e: print("Error:", e)
Error: Prompt must be given after assistant completion, or use `self.cont_pr`.

Setting cont_pr allows a “default prompt” to be specified when a prompt isn’t specified. Usually used to prompt the model to continue.

chat.cont_pr = "keep going..."
chat()

Beyond personal fulfillment, life’s meaning involves: - Making a positive impact on others - Learning and growing continuously - Creating something lasting - Finding love and connection - Discovering your unique purpose - Contributing to something bigger than yourself - Experiencing joy and wonder - Overcoming challenges - Leaving the world better than you found it

  • id: msg_016VF32ScckZzM8MmahBL1Hn
  • content: [{'text': "Beyond personal fulfillment, life's meaning involves:\n- Making a positive impact on others\n- Learning and growing continuously\n- Creating something lasting\n- Finding love and connection\n- Discovering your unique purpose\n- Contributing to something bigger than yourself\n- Experiencing joy and wonder\n- Overcoming challenges\n- Leaving the world better than you found it", 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 105, 'output_tokens': 78, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

We can also use streaming:

chat = Chat(model, sp=sp)
for o in chat("I'm Jeremy", stream=True): print(o, end='')
Hello Jeremy! Nice to meet you. How are you today?
for o in chat(q, prefill=pref, stream=True): print(o, end='')
According to Douglas Adams,  it's 42. More seriously: to find purpose, grow, love, and make a positive impact while experiencing what existence has to offer.

Chat tool use

We automagically get streamlined tool use as well:

pr = f"What is {a}+{b}?"
pr
'What is 604542+6458932?'
chat = Chat(model, sp=sp, tools=[sums])
r = chat(pr)
r
Finding the sum of 604542 and 6458932

Let me calculate that sum for you.

  • id: msg_01GrTXksxD6FuVS4cX6kcqjb
  • content: [{'text': 'Let me calculate that sum for you.', 'type': 'text'}, {'id': 'toolu_01MXtfne5SWH5xJ8d2zu5qnd', 'input': {'a': 604542, 'b': 6458932}, 'name': 'sums', 'type': 'tool_use'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: tool_use
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 437, 'output_tokens': 81, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

Now we need to send this result to Claude—calling the object with no parameters tells it to return the tool result to Claude:

chat()

604542 + 6458932 = 7063474

  • id: msg_018Qd5xBeTvDoYMQHxAj4Am9
  • content: [{'text': '604542 + 6458932 = 7063474', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 532, 'output_tokens': 18, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

It should be correct, because it actually used our Python function to do the addition. Let’s check:

a+b
7063474

Images

Claude can handle image data as well. As everyone knows, when testing image APIs you have to use a cute puppy.

# Image is Cute_dog.jpg from Wikimedia
fn = Path('samples/puppy.jpg')
display.Image(filename=fn, width=200)

img = fn.read_bytes()

Claude expects an image message to have the following structure

{
    'role': 'user', 
    'content': [
        {'type':'text', 'text':'What is in the image?'},
        {
            'type':'image', 
            'source': {
                'type':'base64', 'media_type':'media_type', 'data': 'data'
            }
        }
    ]
}

msglm automatically detects if a message is an image, encodes it, and generates the data structure above. All we need to do is a create a list containing our image and a query and then pass it to mk_msg.

Let’s try it out…

q = "In brief, what color flowers are in this image?"
msg = mk_msg([img, q])
c([msg])

In this adorable puppy photo, there are purple/lavender colored flowers (appears to be asters or similar daisy-like flowers) in the background.

  • id: msg_01MHczh7XATDfQ9uH4kp8dMz
  • content: [{'text': 'In this adorable puppy photo, there are purple/lavender colored flowers (appears to be asters or similar daisy-like flowers) in the background.', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 110, 'output_tokens': 37, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}

You don’t need to call mk_msg on each individual message before passing them to the Chat class. Instead you can pass your messages in a list and the Chat class will automatically call mk_msgs in the background.

c(["How are you?", r])

For messages that contain multiple content types (like an image with a question), you’ll need to enclose the message contents in a list as shown below:

c(["How are you?", r, [img, q]])
c = Chat(model)
c([img, q])

In this adorable puppy photo, there are purple/lavender colored flowers (appears to be asters or similar daisy-like flowers) in the background.

  • id: msg_015jc3XoDD1hCNTYFFy4h8jB
  • content: [{'text': 'In this adorable puppy photo, there are purple/lavender colored flowers (appears to be asters or similar daisy-like flowers) in the background.', 'type': 'text'}]
  • model: claude-3-5-sonnet-20241022
  • role: assistant
  • stop_reason: end_turn
  • stop_sequence: None
  • type: message
  • usage: {'input_tokens': 110, 'output_tokens': 37, 'cache_creation_input_tokens': 0, 'cache_read_input_tokens': 0}
Note

Unfortunately, not all Claude models support images 😞. This table summarizes the capabilities of each Claude model and the different modalities they support.

Caching

Claude supports context caching by adding a cache_control header to the message content.

{
    "role": "user",
    "content": [
        {
            "type": "text", 
            "text": "Please cache my message", 
            "cache_control": {"type": "ephemeral"}
        }
    ]
}

To cache a message, we simply set cache=True when calling mk_msg.

mk_msg(['hi', 'there'], cache=True)
{'role': 'user',
 'content': [{'type': 'text', 'text': 'hi'},
  {'type': 'text', 'text': 'there', 'cache_control': {'type': 'ephemeral'}}]}

Third party providers

Amazon Bedrock

These are Amazon’s current Claude models:

models_aws
['claude-3-5-haiku-20241022',
 'anthropic.claude-3-opus-20240229-v1:0',
 'anthropic.claude-3-5-sonnet-20241022-v2:0',
 'anthropic.claude-3-sonnet-20240229-v1:0']
Note

anthropic at version 0.34.2 seems not to install boto3 as a dependency. You may need to do a pip install boto3 or the creation of the Client below fails.

Provided boto3 is installed, we otherwise don’t need any extra code to support Amazon Bedrock – we just have to set up the approach client:

ab = AnthropicBedrock(
    aws_access_key=os.environ['AWS_ACCESS_KEY'],
    aws_secret_key=os.environ['AWS_SECRET_KEY'],
)
client = Client(models_aws[-1], ab)
chat = Chat(cli=client)
chat("I'm Jeremy")

Google Vertex

models_goog
from anthropic import AnthropicVertex
import google.auth
project_id = google.auth.default()[1]
region = "us-east5"
gv = AnthropicVertex(project_id=project_id, region=region)
client = Client(models_goog[-1], gv)
chat = Chat(cli=client)
chat("I'm Jeremy")