edsl 0.1.39.dev3__py3-none-any.whl → 0.1.39.dev5__py3-none-any.whl

edsl/language_models/key_management/KeyLookupBuilder.py

@@ -0,0 +1,273 @@
+from typing import Optional, List
+from collections import UserDict
+import os
+from functools import lru_cache
+from dataclasses import dataclass, asdict
+
+from edsl.enums import service_to_api_keyname
+from edsl.exceptions.general import MissingAPIKeyError
+
+from edsl.language_models.key_management.KeyLookup import KeyLookup
+
+from edsl.language_models.key_management.models import (
+    APIKeyEntry,
+    LimitEntry,
+    APIIDEntry,
+    LanguageModelInput,
+)
+
+service_to_api_keyname["bedrock"] = "AWS_SECRET_ACCESS_KEY"
+service_to_api_id = {"bedrock": "AWS_ACCESS_KEY_ID"}
+
+api_keyname_to_service = {}
+
+for service, key in service_to_api_keyname.items():
+    if isinstance(key, list):
+        for k in key:
+            api_keyname_to_service[k] = service
+    else:
+        api_keyname_to_service[key] = service
+
+api_id_to_service = {"AWS_ACCESS_KEY_ID": "bedrock"}
+
+
+class KeyLookupBuilder:
+    """Builds KeyLookup options.
+
+    >>> builder = KeyLookupBuilder(fetch_order=("config", "env"))
+    >>> builder.DEFAULT_RPM
+    10
+    >>> builder.DEFAULT_TPM
+    2000000
+    >>> builder.fetch_order
+    ('config', 'env')
+
+    Test invalid fetch_order:
+    >>> try:
+    ...     KeyLookupBuilder(fetch_order=["config", "env"])  # Should be tuple
+    ... except ValueError as e:
+    ...     str(e)
+    'fetch_order must be a tuple'
+
+    Test service extraction:
+    >>> builder.extract_service("EDSL_SERVICE_RPM_OPENAI")
+    ('openai', 'rpm')
+    """
+
+    DEFAULT_RPM = 10
+    DEFAULT_TPM = 2000000
+
+    def __init__(self, fetch_order: Optional[tuple[str]] = None):
+        if fetch_order is None:
+            self.fetch_order = ("config", "env")
+        else:
+            self.fetch_order = fetch_order
+
+        if not isinstance(self.fetch_order, tuple):
+            raise ValueError("fetch_order must be a tuple")
+
+        self.limit_data = {}
+        self.key_data = {}
+        self.id_data = {}
+        self.process_key_value_pairs()
+
+    @property
+    def known_services(self):
+        """Get the set of known services.
+
+        >>> builder = KeyLookupBuilder()
+        >>> isinstance(builder.known_services, set)
+        True
+        """
+        return set(self.key_data.keys()) | set(self.limit_data.keys())
+
+    @lru_cache
+    def build(self) -> "KeyLookup":
+        """Build a KeyLookup instance.
+
+        >>> builder = KeyLookupBuilder()
+        >>> lookup = builder.build()
+        >>> isinstance(lookup, KeyLookup)
+        True
+        >>> lookup['test'].api_token  # Test service should always exist
+        'test'
+        """
+        d = {}
+        for service in self.known_services:
+            try:
+                d[service] = self.get_language_model_input(service)
+            except MissingAPIKeyError:
+                pass
+
+        d.update({"test": LanguageModelInput(api_token="test", rpm=10, tpm=2000000)})
+        return KeyLookup(d)
+
+    def get_language_model_input(self, service: str) -> LanguageModelInput:
+        """Get the language model input for a given service.
+
+        >>> builder = KeyLookupBuilder()
+        >>> try:
+        ...     builder.get_language_model_input("nonexistent_service")
+        ... except MissingAPIKeyError as e:
+        ...     str(e)
+        "No key found for service 'nonexistent_service'"
+        """
+        if (key_entries := self.key_data.get(service)) is None:
+            raise MissingAPIKeyError(f"No key found for service '{service}'")
+
+        if len(key_entries) == 1:
+            api_key_entry = key_entries[0]
+
+        id_entry = self.id_data.get(service)
+        id_source = id_entry.source if id_entry is not None else None
+        api_id = id_entry.value if id_entry is not None else None
+
+        if (limit_entry := self.limit_data.get(service)) is None:
+            limit_entry = LimitEntry(
+                service=service,
+                rpm=self.DEFAULT_RPM,
+                tpm=self.DEFAULT_TPM,
+                source="default",
+            )
+
+        if limit_entry.rpm is None:
+            limit_entry.rpm = self.DEFAULT_RPM
+        if limit_entry.tpm is None:
+            limit_entry.tpm = self.DEFAULT_TPM
+
+        return LanguageModelInput(
+            api_token=api_key_entry.value,
+            rpm=int(limit_entry.rpm),
+            tpm=int(limit_entry.tpm),
+            api_id=api_id,
+            token_source=api_key_entry.source,
+            limit_source=limit_entry.source,
+            id_source=id_source,
+        )
+
+    def __repr__(self):
+        return f"DataSource(key_data={self.key_data}, limit_data={self.limit_data}, id_data={self.id_data})"
+
+    def _os_env_key_value_pairs(self):
+        return dict(list(os.environ.items()))
+
+    def _coop_key_value_pairs(self):
+        from edsl.coop import Coop
+
+        c = Coop()
+        return dict(list(c.fetch_rate_limit_config_vars().items()))
+
+    def _config_key_value_pairs(self):
+        from edsl.config import CONFIG
+
+        return dict(list(CONFIG.items()))
+
+    @staticmethod
+    def extract_service(key: str) -> str:
+        """Extract the service and limit type from the key"""
+        limit_type, service_raw = key.replace("EDSL_SERVICE_", "").split("_")
+        return service_raw.lower(), limit_type.lower()
+
+    def get_key_value_pairs(self) -> dict:
+        """Get key-value pairs from configured sources."""
+        fetching_functions = {
+            "env": self._os_env_key_value_pairs,
+            "coop": self._coop_key_value_pairs,
+            "config": self._config_key_value_pairs,
+        }
+        d = {}
+        for source in self.fetch_order:
+            f = fetching_functions[source]
+            new_data = f()
+            for k, v in new_data.items():
+                d[k] = (v, source)
+        return d
+
+    def _entry_type(self, key, value) -> str:
+        """Determine the type of entry from a key.
+
+        >>> builder = KeyLookupBuilder()
+        >>> builder._entry_type("EDSL_SERVICE_RPM_OPENAI", "60")
+        'limit'
+        >>> builder._entry_type("OPENAI_API_KEY", "sk-1234")
+        'api_key'
+        >>> builder._entry_type("AWS_ACCESS_KEY_ID", "AKIA1234")
+        'api_id'
+        >>> builder._entry_type("UNKNOWN_KEY", "value")
+        'unknown'
+        """
+        if key.startswith("EDSL_SERVICE_"):
+            return "limit"
+        elif key in api_keyname_to_service:
+            return "api_key"
+        elif key in api_id_to_service:
+            return "api_id"
+        return "unknown"
+
+    def _add_id(self, key: str, value: str, source: str) -> None:
+        """Add an API ID to the id_data dictionary.
+
+        >>> builder = KeyLookupBuilder()
+        >>> builder._add_id("AWS_ACCESS_KEY_ID", "AKIA1234", "env")
+        >>> builder.id_data["bedrock"].value
+        'AKIA1234'
+        >>> try:
+        ...     builder._add_id("AWS_ACCESS_KEY_ID", "AKIA5678", "env")
+        ... except ValueError as e:
+        ...     str(e)
+        'Duplicate ID for service bedrock'
+        """
+        service = api_id_to_service[key]
+        if service not in self.id_data:
+            self.id_data[service] = APIIDEntry(
+                service=service, name=key, value=value, source=source
+            )
+        else:
+            raise ValueError(f"Duplicate ID for service {service}")
+
+    def _add_limit(self, key: str, value: str, source: str) -> None:
+        """Add a rate limit entry to the limit_data dictionary.
+
+        >>> builder = KeyLookupBuilder()
+        >>> builder._add_limit("EDSL_SERVICE_RPM_OPENAI", "60", "config")
+        >>> builder.limit_data["openai"].rpm
+        '60'
+        >>> builder._add_limit("EDSL_SERVICE_TPM_OPENAI", "100000", "config")
+        >>> builder.limit_data["openai"].tpm
+        '100000'
+        """
+        service, limit_type = self.extract_service(key)
+        if service in self.limit_data:
+            setattr(self.limit_data[service], limit_type.lower(), value)
+        else:
+            new_limit_entry = LimitEntry(
+                service=service, rpm=None, tpm=None, source=source
+            )
+            setattr(new_limit_entry, limit_type.lower(), value)
+            self.limit_data[service] = new_limit_entry
+
+    def _add_api_key(self, key: str, value: str, source: str) -> None:
+        """Add an API key entry to the key_data dictionary.
+
+        >>> builder = KeyLookupBuilder()
+        >>> builder._add_api_key("OPENAI_API_KEY", "sk-1234", "env")
+        >>> 'sk-1234' == builder.key_data["openai"][-1].value
+        True
+        """
+        service = api_keyname_to_service[key]
+        new_entry = APIKeyEntry(service=service, name=key, value=value, source=source)
+        if service not in self.key_data:
+            self.key_data[service] = [new_entry]
+        else:
+            self.key_data[service].append(new_entry)
+
+    def process_key_value_pairs(self) -> None:
+        """Process all key-value pairs from the configured sources."""
+        for key, value_pair in self.get_key_value_pairs().items():
+            value, source = value_pair
+            if (entry_type := self._entry_type(key, value)) == "limit":
+                self._add_limit(key, value, source)
+            elif entry_type == "api_key":
+                self._add_api_key(key, value, source)
+            elif entry_type == "api_id":
+                self._add_id(key, value, source)

edsl/language_models/key_management/KeyLookupCollection.py

@@ -0,0 +1,38 @@
+from collections import UserDict
+
+from edsl.language_models.key_management.KeyLookupBuilder import KeyLookupBuilder
+
+
+class KeyLookupCollection(UserDict):
+    """A singleton class that stores key-lookup objects.
+
+    This is because once a KeyLook is created once, we do not
+    need to keep re-creating it.
+
+    >>> collection = KeyLookupCollection()
+    >>> collection2 = KeyLookupCollection()
+    >>> collection is collection2  # Test singleton pattern
+    True
+    >>> collection.add_key_lookup(("config", "env"))
+    >>> ("config", "env") in collection.data
+    True
+    """
+
+    _instance = None
+
+    def __new__(cls, *args, **kwargs):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self, *args, **kwargs):
+        if not hasattr(self, "_initialized"):
+            self.data = {}
+            self._initialized = True
+            super().__init__(*args, **kwargs)
+
+    def add_key_lookup(self, fetch_order=None):
+        if fetch_order is None:
+            fetch_order = ("config", "env")
+        if fetch_order not in self.data:
+            self.data[fetch_order] = KeyLookupBuilder(fetch_order=fetch_order).build()

edsl/language_models/key_management/models.py

@@ -0,0 +1,131 @@
+from dataclasses import dataclass, asdict
+from typing import Optional
+
+
+@dataclass
+class APIKeyEntry:
+    """A class representing an API key entry.
+
+    >>> entry = APIKeyEntry.example()
+    >>> entry.service
+    'openai'
+    >>> entry.name
+    'OPENAI_API_KEY'
+    >>> entry.value
+    'sk-abcd1234'
+    >>> entry.source
+    'env'
+    """
+
+    service: str
+    name: str
+    value: str
+    source: Optional[str] = None
+
+    @classmethod
+    def example(cls):
+        return APIKeyEntry(
+            service="openai", name="OPENAI_API_KEY", value="sk-abcd1234", source="env"
+        )
+
+
+@dataclass
+class LimitEntry:
+    """A class representing rate limit entries for a service.
+
+    >>> limit = LimitEntry.example()
+    >>> limit.service
+    'openai'
+    >>> limit.rpm
+    60
+    >>> limit.tpm
+    100000
+    >>> limit.source
+    'config'
+    """
+
+    service: str
+    rpm: int
+    tpm: int
+    source: Optional[str] = None
+
+    @classmethod
+    def example(cls):
+        return LimitEntry(service="openai", rpm=60, tpm=100000, source="config")
+
+
+@dataclass
+class APIIDEntry:
+    """A class representing an API ID entry.
+
+    >>> id_entry = APIIDEntry.example()
+    >>> id_entry.service
+    'bedrock'
+    >>> id_entry.name
+    'AWS_ACCESS_KEY_ID'
+    >>> id_entry.value
+    'AKIA1234'
+    >>> id_entry.source
+    'env'
+    """
+
+    service: str
+    name: str
+    value: str
+    source: Optional[str] = None
+
+    @classmethod
+    def example(cls):
+        return APIIDEntry(
+            service="bedrock", name="AWS_ACCESS_KEY_ID", value="AKIA1234", source="env"
+        )
+
+
+@dataclass
+class LanguageModelInput:
+    """A class representing input configuration for a language model service.
+
+    >>> lm_input = LanguageModelInput.example()
+    >>> lm_input.api_token
+    'sk-abcd123'
+    >>> lm_input.rpm
+    60
+    >>> lm_input.tpm
+    100000
+    >>> lm_input.api_id
+
+
+    Test dictionary conversion:
+    >>> d = lm_input.to_dict()
+    >>> isinstance(d, dict)
+    True
+    >>> LanguageModelInput.from_dict(d).api_token == lm_input.api_token
+    True
+    """
+
+    api_token: str
+    rpm: int
+    tpm: int
+    api_id: Optional[str] = None
+    token_source: Optional[str] = None
+    limit_source: Optional[str] = None
+    id_source: Optional[str] = None
+
+    def to_dict(self):
+        return asdict(self)
+
+    @classmethod
+    def from_dict(cls, d):
+        return cls(**d)
+
+    @classmethod
+    def example(cls):
+        return LanguageModelInput(
+            api_token="sk-abcd123", tpm=100000, rpm=60, api_id=None
+        )
+
+
+if __name__ == "__main__":
+    import doctest
+
+    doctest.testmod()

edsl/language_models/model.py

@@ -0,0 +1,256 @@
+import textwrap
+from random import random
+from typing import Optional, TYPE_CHECKING, List
+
+from edsl.utilities.PrettyList import PrettyList
+from edsl.config import CONFIG
+
+from edsl.inference_services.InferenceServicesCollection import (
+    InferenceServicesCollection,
+)
+from edsl.inference_services.data_structures import AvailableModels
+from edsl.inference_services.InferenceServiceABC import InferenceServiceABC
+from edsl.enums import InferenceServiceLiteral
+
+if TYPE_CHECKING:
+    from edsl.results.Dataset import Dataset
+
+
+def get_model_class(model_name, registry: Optional[InferenceServicesCollection] = None):
+    from edsl.inference_services.registry import default
+
+    registry = registry or default
+    factory = registry.create_model_factory(model_name)
+    return factory
+
+
+class Meta(type):
+    def __repr__(cls):
+        return textwrap.dedent(
+            f"""\
+        Available models: {cls.available()}
+        
+        To create an instance, you can do: 
+        >>> m = Model('gpt-4-1106-preview', temperature=0.5, ...)
+        
+        To get the default model, you can leave out the model name. 
+        To see the available models, you can do:
+        >>> Model.available()
+        """
+        )
+
+
+class Model(metaclass=Meta):
+    default_model = CONFIG.get("EDSL_DEFAULT_MODEL")
+    _registry: InferenceServicesCollection = None  # Class-level registry storage
+
+    @classmethod
+    def get_registry(cls) -> InferenceServicesCollection:
+        """Get the current registry or initialize with default if None"""
+        if cls._registry is None:
+            from edsl.inference_services.registry import default
+
+            cls._registry = default
+        return cls._registry
+
+    @classmethod
+    def set_registry(cls, registry: InferenceServicesCollection) -> None:
+        """Set a new registry"""
+        cls._registry = registry
+
+    def __new__(
+        cls,
+        model_name: Optional[str] = None,
+        service_name: Optional[InferenceServiceLiteral] = None,
+        registry: Optional[InferenceServicesCollection] = None,
+        *args,
+        **kwargs,
+    ):
+        "Instantiate a new language model."
+        # Map index to the respective subclass
+        if model_name is None:
+            model_name = (
+                cls.default_model
+            )  # when model_name is None, use the default model, set in the config file
+
+        if registry is not None:
+            cls.set_registry(registry)
+
+        if isinstance(model_name, int):  # can refer to a model by index
+            model_name = cls.available(name_only=True)[model_name]
+
+        factory = cls.get_registry().create_model_factory(
+            model_name, service_name=service_name
+        )
+        return factory(*args, **kwargs)
+
+    @classmethod
+    def add_model(cls, service_name, model_name) -> None:
+        cls.get_registry().add_model(service_name, model_name)
+
+    @classmethod
+    def service_classes(cls) -> List["InferenceServiceABC"]:
+        """Returns a list of service classes.
+
+        >>> Model.service_classes()
+        [...]
+        """
+        return [r for r in cls.services(name_only=True)]
+
+    @classmethod
+    def services(cls, name_only: bool = False) -> List[str]:
+        """Returns a list of services, annotated with whether the user has local keys for them."""
+        services_with_local_keys = set(cls.key_info().select("service").to_list())
+        f = lambda service_name: (
+            "yes" if service_name in services_with_local_keys else " "
+        )
+        if name_only:
+            return PrettyList(
+                [r._inference_service_ for r in cls.get_registry().services],
+                columns=["Service Name"],
+            )
+        else:
+            return PrettyList(
+                [
+                    (r._inference_service_, f(r._inference_service_))
+                    for r in cls.get_registry().services
+                ],
+                columns=["Service Name", "Local key?"],
+            )
+
+    @classmethod
+    def services_with_local_keys(cls) -> set:
+        """Returns a list of services for which the user has local keys."""
+        return set(cls.key_info().select("service").to_list())
+
+    @classmethod
+    def key_info(cls, obscure_api_key: bool = True) -> "Dataset":
+        """Returns a dataset of local key information."""
+        from edsl.language_models.key_management.KeyLookupCollection import (
+            KeyLookupCollection,
+        )
+        from edsl.scenarios import Scenario, ScenarioList
+
+        klc = KeyLookupCollection()
+        klc.add_key_lookup(fetch_order=None)
+        sl = ScenarioList()
+        for service, entry in list(klc.data.values())[0].items():
+            sl.append(Scenario({"service": service} | entry.to_dict()))
+        if obscure_api_key:
+            for service in sl:
+                service["api_token"] = (
+                    service["api_token"][:4] + "..." + service["api_token"][-4:]
+                )
+        return sl.to_dataset()
+
+    @classmethod
+    def search_models(cls, search_term: str):
+        return cls.available(search_term=search_term)
+
+    @classmethod
+    def all_known_models(cls) -> "AvailableModels":
+        return cls.get_registry().available()
+
+    @classmethod
+    def available_with_local_keys(cls):
+        services_with_local_keys = set(cls.key_info().select("service").to_list())
+        return [
+            m
+            for m in cls.get_registry().available()
+            if m.service_name in services_with_local_keys
+        ]
+
+    @classmethod
+    def available(
+        cls,
+        search_term: str = None,
+        name_only: bool = False,
+        service: Optional[str] = None,
+    ):
+        # if search_term is None and service is None:
+        #     print("Getting available models...")
+        #     print("You have local keys for the following services:")
+        #     print(cls.services_with_local_keys())
+        #     print("\n")
+        #     print("To see models by service, use the 'service' parameter.")
+        #     print("E.g., Model.available(service='openai')")
+        #     return None
+
+        if service is not None:
+            if service not in cls.services(name_only=True):
+                raise ValueError(
+                    f"Service {service} not found in available services.",
+                    f"Available services are: {cls.services()}",
+                )
+
+        full_list = cls.get_registry().available(service=service)
+
+        if search_term is None:
+            if name_only:
+                return PrettyList(
+                    [m.model_name for m in full_list],
+                    columns=["Model Name"],
+                )
+            else:
+                return PrettyList(
+                    [[m.model_name, m.service_name] for m in full_list],
+                    columns=["Model Name", "Service Name"],
+                )
+        else:
+            filtered_results = [
+                m
+                for m in full_list
+                if search_term in m.model_name or search_term in m.service_name
+            ]
+            if name_only:
+                return PrettyList(
+                    [m.model_name for m in filtered_results],
+                    columns=["Model Name"],
+                )
+            else:
+                return PrettyList(
+                    [[m.model_name, m.service_name] for m in full_list],
+                    columns=["Model Name", "Service Name"],
+                )
+
+    @classmethod
+    def check_models(cls, verbose=False):
+        print("Checking all available models...\n")
+        for model in cls.available(name_only=True):
+            print(f"Now checking: {model}")
+            try:
+                m = cls(model)
+            except Exception as e:
+                print(f"Error creating instance of {model}: {e}")
+                continue
+            try:
+                results = m.hello(verbose)
+                if verbose:
+                    print(f"Results from model call: {results}")
+            except Exception as e:
+                print(f"Error calling 'hello' on {model}: {e}")
+                continue
+            print("OK!")
+            print("\n")
+
+    @classmethod
+    def example(cls, randomize: bool = False) -> "Model":
+        """
+        Returns an example Model instance.
+
+        :param randomize: If True, the temperature is set to a random decimal between 0 and 1.
+        """
+        temperature = 0.5 if not randomize else round(random(), 2)
+        model_name = cls.default_model
+        return cls(model_name, temperature=temperature)
+
+
+if __name__ == "__main__":
+    import doctest
+
+    doctest.testmod(optionflags=doctest.ELLIPSIS)
+
+    available = Model.available()
+    m = Model("gpt-4-1106-preview")
+    results = m.execute_model_call("Hello world")
+    print(results)