from typing import Any, Dict, List, Optional, Sequence
from langchain_core.callbacks import CallbackManagerForLLMRun
from langchain_core.language_models.llms import LLM
from langchain_core.utils import convert_to_secret_str, get_from_dict_or_env, pre_init
from pydantic import ConfigDict, SecretStr
from langchain_community.llms.utils import enforce_stop_tokens
[docs]
class AlephAlpha(LLM):
"""Aleph Alpha large language models.
To use, you should have the ``aleph_alpha_client`` python package installed, and the
environment variable ``ALEPH_ALPHA_API_KEY`` set with your API key, or pass
it as a named parameter to the constructor.
Parameters are explained more in depth here:
https://github.com/Aleph-Alpha/aleph-alpha-client/blob/c14b7dd2b4325c7da0d6a119f6e76385800e097b/aleph_alpha_client/completion.py#L10
Example:
.. code-block:: python
from langchain_community.llms import AlephAlpha
aleph_alpha = AlephAlpha(aleph_alpha_api_key="my-api-key")
"""
client: Any = None #: :meta private:
model: Optional[str] = "luminous-base"
"""Model name to use."""
maximum_tokens: int = 64
"""The maximum number of tokens to be generated."""
temperature: float = 0.0
"""A non-negative float that tunes the degree of randomness in generation."""
top_k: int = 0
"""Number of most likely tokens to consider at each step."""
top_p: float = 0.0
"""Total probability mass of tokens to consider at each step."""
presence_penalty: float = 0.0
"""Penalizes repeated tokens."""
frequency_penalty: float = 0.0
"""Penalizes repeated tokens according to frequency."""
repetition_penalties_include_prompt: Optional[bool] = False
"""Flag deciding whether presence penalty or frequency penalty are
updated from the prompt."""
use_multiplicative_presence_penalty: Optional[bool] = False
"""Flag deciding whether presence penalty is applied
multiplicatively (True) or additively (False)."""
penalty_bias: Optional[str] = None
"""Penalty bias for the completion."""
penalty_exceptions: Optional[List[str]] = None
"""List of strings that may be generated without penalty,
regardless of other penalty settings"""
penalty_exceptions_include_stop_sequences: Optional[bool] = None
"""Should stop_sequences be included in penalty_exceptions."""
best_of: Optional[int] = None
"""returns the one with the "best of" results
(highest log probability per token)
"""
n: int = 1
"""How many completions to generate for each prompt."""
logit_bias: Optional[Dict[int, float]] = None
"""The logit bias allows to influence the likelihood of generating tokens."""
log_probs: Optional[int] = None
"""Number of top log probabilities to be returned for each generated token."""
tokens: Optional[bool] = False
"""return tokens of completion."""
disable_optimizations: Optional[bool] = False
minimum_tokens: Optional[int] = 0
"""Generate at least this number of tokens."""
echo: bool = False
"""Echo the prompt in the completion."""
use_multiplicative_frequency_penalty: bool = False
sequence_penalty: float = 0.0
sequence_penalty_min_length: int = 2
use_multiplicative_sequence_penalty: bool = False
completion_bias_inclusion: Optional[Sequence[str]] = None
completion_bias_inclusion_first_token_only: bool = False
completion_bias_exclusion: Optional[Sequence[str]] = None
completion_bias_exclusion_first_token_only: bool = False
"""Only consider the first token for the completion_bias_exclusion."""
contextual_control_threshold: Optional[float] = None
"""If set to None, attention control parameters only apply to those tokens that have
explicitly been set in the request.
If set to a non-None value, control parameters are also applied to similar tokens.
"""
control_log_additive: Optional[bool] = True
"""True: apply control by adding the log(control_factor) to attention scores.
False: (attention_scores - - attention_scores.min(-1)) * control_factor
"""
repetition_penalties_include_completion: bool = True
"""Flag deciding whether presence penalty or frequency penalty
are updated from the completion."""
raw_completion: bool = False
"""Force the raw completion of the model to be returned."""
stop_sequences: Optional[List[str]] = None
"""Stop sequences to use."""
# Client params
aleph_alpha_api_key: Optional[SecretStr] = None
"""API key for Aleph Alpha API."""
host: str = "https://api.aleph-alpha.com"
"""The hostname of the API host.
The default one is "https://api.aleph-alpha.com")"""
hosting: Optional[str] = None
"""Determines in which datacenters the request may be processed.
You can either set the parameter to "aleph-alpha" or omit it (defaulting to None).
Not setting this value, or setting it to None, gives us maximal
flexibility in processing your request in our
own datacenters and on servers hosted with other providers.
Choose this option for maximal availability.
Setting it to "aleph-alpha" allows us to only process the
request in our own datacenters.
Choose this option for maximal data privacy."""
request_timeout_seconds: int = 305
"""Client timeout that will be set for HTTP requests in the
`requests` library's API calls.
Server will close all requests after 300 seconds with an internal server error."""
total_retries: int = 8
"""The number of retries made in case requests fail with certain retryable
status codes. If the last
retry fails a corresponding exception is raised. Note, that between retries
an exponential backoff
is applied, starting with 0.5 s after the first retry and doubling for
each retry made. So with the
default setting of 8 retries a total wait time of 63.5 s is added
between the retries."""
nice: bool = False
"""Setting this to True, will signal to the API that you intend to be
nice to other users
by de-prioritizing your request below concurrent ones."""
model_config = ConfigDict(
extra="forbid",
)
[docs]
@pre_init
def validate_environment(cls, values: Dict) -> Dict:
"""Validate that api key and python package exists in environment."""
values["aleph_alpha_api_key"] = convert_to_secret_str(
get_from_dict_or_env(values, "aleph_alpha_api_key", "ALEPH_ALPHA_API_KEY")
)
try:
from aleph_alpha_client import Client
values["client"] = Client(
token=values["aleph_alpha_api_key"].get_secret_value(),
host=values["host"],
hosting=values["hosting"],
request_timeout_seconds=values["request_timeout_seconds"],
total_retries=values["total_retries"],
nice=values["nice"],
)
except ImportError:
raise ImportError(
"Could not import aleph_alpha_client python package. "
"Please install it with `pip install aleph_alpha_client`."
)
return values
@property
def _default_params(self) -> Dict[str, Any]:
"""Get the default parameters for calling the Aleph Alpha API."""
return {
"maximum_tokens": self.maximum_tokens,
"temperature": self.temperature,
"top_k": self.top_k,
"top_p": self.top_p,
"presence_penalty": self.presence_penalty,
"frequency_penalty": self.frequency_penalty,
"n": self.n,
"repetition_penalties_include_prompt": self.repetition_penalties_include_prompt, # noqa: E501
"use_multiplicative_presence_penalty": self.use_multiplicative_presence_penalty, # noqa: E501
"penalty_bias": self.penalty_bias,
"penalty_exceptions": self.penalty_exceptions,
"penalty_exceptions_include_stop_sequences": self.penalty_exceptions_include_stop_sequences, # noqa: E501
"best_of": self.best_of,
"logit_bias": self.logit_bias,
"log_probs": self.log_probs,
"tokens": self.tokens,
"disable_optimizations": self.disable_optimizations,
"minimum_tokens": self.minimum_tokens,
"echo": self.echo,
"use_multiplicative_frequency_penalty": self.use_multiplicative_frequency_penalty, # noqa: E501
"sequence_penalty": self.sequence_penalty,
"sequence_penalty_min_length": self.sequence_penalty_min_length,
"use_multiplicative_sequence_penalty": self.use_multiplicative_sequence_penalty, # noqa: E501
"completion_bias_inclusion": self.completion_bias_inclusion,
"completion_bias_inclusion_first_token_only": self.completion_bias_inclusion_first_token_only, # noqa: E501
"completion_bias_exclusion": self.completion_bias_exclusion,
"completion_bias_exclusion_first_token_only": self.completion_bias_exclusion_first_token_only, # noqa: E501
"contextual_control_threshold": self.contextual_control_threshold,
"control_log_additive": self.control_log_additive,
"repetition_penalties_include_completion": self.repetition_penalties_include_completion, # noqa: E501
"raw_completion": self.raw_completion,
}
@property
def _identifying_params(self) -> Dict[str, Any]:
"""Get the identifying parameters."""
return {**{"model": self.model}, **self._default_params}
@property
def _llm_type(self) -> str:
"""Return type of llm."""
return "aleph_alpha"
def _call(
self,
prompt: str,
stop: Optional[List[str]] = None,
run_manager: Optional[CallbackManagerForLLMRun] = None,
**kwargs: Any,
) -> str:
"""Call out to Aleph Alpha's completion endpoint.
Args:
prompt: The prompt to pass into the model.
stop: Optional list of stop words to use when generating.
Returns:
The string generated by the model.
Example:
.. code-block:: python
response = aleph_alpha("Tell me a joke.")
"""
from aleph_alpha_client import CompletionRequest, Prompt
params = self._default_params
if self.stop_sequences is not None and stop is not None:
raise ValueError(
"stop sequences found in both the input and default params."
)
elif self.stop_sequences is not None:
params["stop_sequences"] = self.stop_sequences
else:
params["stop_sequences"] = stop
params = {**params, **kwargs}
request = CompletionRequest(prompt=Prompt.from_text(prompt), **params)
response = self.client.complete(model=self.model, request=request)
text = response.completions[0].completion
# If stop tokens are provided, Aleph Alpha's endpoint returns them.
# In order to make this consistent with other endpoints, we strip them.
if stop is not None or self.stop_sequences is not None:
text = enforce_stop_tokens(text, params["stop_sequences"])
return text
if __name__ == "__main__":
aa = AlephAlpha() # type: ignore[call-arg]
print(aa.invoke("How are you?")) # noqa: T201
