letta-nightly 0.4.1.dev20241004012408__py3-none-any.whl → 0.4.1.dev20241005104008__py3-none-any.whl

letta/llm_api/llm_api_tools.py

@@ -9,7 +9,6 @@ from typing import List, Optional, Union
 import requests
 
 from letta.constants import CLI_WARNING_PREFIX, OPENAI_CONTEXT_WINDOW_ERROR_SUBSTRING
-from letta.credentials import LettaCredentials
 from letta.llm_api.anthropic import anthropic_chat_completions_request
 from letta.llm_api.azure_openai import (
     MODEL_TO_AZURE_ENGINE,
@@ -29,6 +28,7 @@ from letta.local_llm.constants import (
     INNER_THOUGHTS_KWARG,
     INNER_THOUGHTS_KWARG_DESCRIPTION,
 )
+from letta.providers import GoogleAIProvider
 from letta.schemas.enums import OptionState
 from letta.schemas.llm_config import LLMConfig
 from letta.schemas.message import Message
@@ -37,14 +37,14 @@ from letta.schemas.openai.chat_completion_request import (
     Tool,
     cast_message_to_subtype,
 )
-from letta.schemas.openai.chat_completion_response import ChatCompletionResponse
+from letta.schemas.openai.chat_completion_response import ChatCompletionResponse, Choice
 from letta.streaming_interface import (
     AgentChunkStreamingInterface,
     AgentRefreshStreamingInterface,
 )
 from letta.utils import json_dumps
 
-LLM_API_PROVIDER_OPTIONS = ["openai", "azure", "anthropic", "google_ai", "cohere", "local"]
+LLM_API_PROVIDER_OPTIONS = ["openai", "azure", "anthropic", "google_ai", "cohere", "local", "groq"]
 
 
 # TODO update to use better types
@@ -83,7 +83,7 @@ def add_inner_thoughts_to_functions(
     return new_functions
 
 
-def unpack_inner_thoughts_from_kwargs(
+def unpack_all_inner_thoughts_from_kwargs(
     response: ChatCompletionResponse,
     inner_thoughts_key: str,
 ) -> ChatCompletionResponse:
@@ -93,36 +93,7 @@ def unpack_inner_thoughts_from_kwargs(
 
     new_choices = []
     for choice in response.choices:
-        msg = choice.message
-        if msg.role == "assistant" and msg.tool_calls and len(msg.tool_calls) >= 1:
-            if len(msg.tool_calls) > 1:
-                warnings.warn(f"Unpacking inner thoughts from more than one tool call ({len(msg.tool_calls)}) is not supported")
-            # TODO support multiple tool calls
-            tool_call = msg.tool_calls[0]
-
-            try:
-                # Sadly we need to parse the JSON since args are in string format
-                func_args = dict(json.loads(tool_call.function.arguments))
-                if inner_thoughts_key in func_args:
-                    # extract the inner thoughts
-                    inner_thoughts = func_args.pop(inner_thoughts_key)
-
-                    # replace the kwargs
-                    new_choice = choice.model_copy(deep=True)
-                    new_choice.message.tool_calls[0].function.arguments = json_dumps(func_args)
-                    # also replace the message content
-                    if new_choice.message.content is not None:
-                        warnings.warn(f"Overwriting existing inner monologue ({new_choice.message.content}) with kwarg ({inner_thoughts})")
-                    new_choice.message.content = inner_thoughts
-
-                    # save copy
-                    new_choices.append(new_choice)
-                else:
-                    warnings.warn(f"Did not find inner thoughts in tool call: {str(tool_call)}")
-
-            except json.JSONDecodeError as e:
-                warnings.warn(f"Failed to strip inner thoughts from kwargs: {e}")
-                raise e
+        new_choices.append(unpack_inner_thoughts_from_kwargs(choice, inner_thoughts_key))
 
     # return an updated copy
     new_response = response.model_copy(deep=True)
@@ -130,6 +101,38 @@ def unpack_inner_thoughts_from_kwargs(
     return new_response
 
 
+def unpack_inner_thoughts_from_kwargs(choice: Choice, inner_thoughts_key: str) -> Choice:
+    message = choice.message
+    if message.role == "assistant" and message.tool_calls and len(message.tool_calls) >= 1:
+        if len(message.tool_calls) > 1:
+            warnings.warn(f"Unpacking inner thoughts from more than one tool call ({len(message.tool_calls)}) is not supported")
+        # TODO support multiple tool calls
+        tool_call = message.tool_calls[0]
+
+        try:
+            # Sadly we need to parse the JSON since args are in string format
+            func_args = dict(json.loads(tool_call.function.arguments))
+            if inner_thoughts_key in func_args:
+                # extract the inner thoughts
+                inner_thoughts = func_args.pop(inner_thoughts_key)
+
+                # replace the kwargs
+                new_choice = choice.model_copy(deep=True)
+                new_choice.message.tool_calls[0].function.arguments = json_dumps(func_args)
+                # also replace the message content
+                if new_choice.message.content is not None:
+                    warnings.warn(f"Overwriting existing inner monologue ({new_choice.message.content}) with kwarg ({inner_thoughts})")
+                new_choice.message.content = inner_thoughts
+
+                return new_choice
+            else:
+                warnings.warn(f"Did not find inner thoughts in tool call: {str(tool_call)}")
+
+        except json.JSONDecodeError as e:
+            warnings.warn(f"Failed to strip inner thoughts from kwargs: {e}")
+            raise e
+
+
 def is_context_overflow_error(exception: requests.exceptions.RequestException) -> bool:
     """Checks if an exception is due to context overflow (based on common OpenAI response messages)"""
     from letta.utils import printd
@@ -246,15 +249,17 @@ def create(
     # TODO move to llm_config?
     # if unspecified (None), default to something we've tested
     inner_thoughts_in_kwargs: OptionState = OptionState.DEFAULT,
+    model_settings: Optional[dict] = None,  # TODO: eventually pass from server
 ) -> ChatCompletionResponse:
     """Return response to chat completion with backoff"""
     from letta.utils import printd
 
-    printd(f"Using model {llm_config.model_endpoint_type}, endpoint: {llm_config.model_endpoint}")
+    if not model_settings:
+        from letta.settings import model_settings
 
-    # TODO eventually refactor so that credentials are passed through
+        model_settings = model_settings
 
-    credentials = LettaCredentials.load()
+    printd(f"Using model {llm_config.model_endpoint_type}, endpoint: {llm_config.model_endpoint}")
 
     if function_call and not functions:
         printd("unsetting function_call because functions is None")
@@ -286,7 +291,7 @@ def create(
         ]
 
         # TODO do the same for Azure?
-        if credentials.openai_key is None and llm_config.model_endpoint == "https://api.openai.com/v1":
+        if model_settings.openai_api_key is None and llm_config.model_endpoint == "https://api.openai.com/v1":
             # only is a problem if we are *not* using an openai proxy
             raise ValueError(f"OpenAI key is missing from letta config file")
         if use_tool_naming:
@@ -323,7 +328,7 @@ def create(
             ), type(stream_inferface)
             response = openai_chat_completions_process_stream(
                 url=llm_config.model_endpoint,  # https://api.openai.com/v1 -> https://api.openai.com/v1/chat/completions
-                api_key=credentials.openai_key,
+                api_key=model_settings.openai_api_key,
                 chat_completion_request=data,
                 stream_inferface=stream_inferface,
             )
@@ -332,10 +337,9 @@ def create(
             if isinstance(stream_inferface, AgentChunkStreamingInterface):
                 stream_inferface.stream_start()
             try:
-
                 response = openai_chat_completions_request(
                     url=llm_config.model_endpoint,  # https://api.openai.com/v1 -> https://api.openai.com/v1/chat/completions
-                    api_key=credentials.openai_key,
+                    api_key=model_settings.openai_api_key,
                     chat_completion_request=data,
                 )
             finally:
@@ -343,7 +347,7 @@ def create(
                     stream_inferface.stream_end()
 
         if inner_thoughts_in_kwargs:
-            response = unpack_inner_thoughts_from_kwargs(response=response, inner_thoughts_key=INNER_THOUGHTS_KWARG)
+            response = unpack_all_inner_thoughts_from_kwargs(response=response, inner_thoughts_key=INNER_THOUGHTS_KWARG)
 
         return response
 
@@ -353,7 +357,7 @@ def create(
             raise NotImplementedError(f"Streaming not yet implemented for {llm_config.model_endpoint_type}")
 
         azure_deployment = (
-            credentials.azure_deployment if credentials.azure_deployment is not None else MODEL_TO_AZURE_ENGINE[llm_config.model]
+            model_settings.azure_deployment if model_settings.azure_deployment is not None else MODEL_TO_AZURE_ENGINE[llm_config.model]
         )
         if use_tool_naming:
             data = dict(
@@ -374,10 +378,10 @@ def create(
                 user=str(user_id),
             )
         return azure_openai_chat_completions_request(
-            resource_name=credentials.azure_endpoint,
+            resource_name=model_settings.azure_endpoint,
             deployment_id=azure_deployment,
-            api_version=credentials.azure_version,
-            api_key=credentials.azure_key,
+            api_version=model_settings.azure_version,
+            api_key=model_settings.azure_key,
             data=data,
         )
 
@@ -400,9 +404,9 @@ def create(
 
         return google_ai_chat_completions_request(
             inner_thoughts_in_kwargs=google_ai_inner_thoughts_in_kwarg,
-            service_endpoint=credentials.google_ai_service_endpoint,
+            service_endpoint=GoogleAIProvider(model_settings.gemini_api_key).service_endpoint,
             model=llm_config.model,
-            api_key=credentials.google_ai_key,
+            api_key=model_settings.gemini_api_key,
             # see structure of payload here: https://ai.google.dev/docs/function_calling
             data=dict(
                 contents=[m.to_google_ai_dict() for m in messages],
@@ -424,7 +428,7 @@ def create(
 
         return anthropic_chat_completions_request(
             url=llm_config.model_endpoint,
-            api_key=credentials.anthropic_key,
+            api_key=model_settings.anthropic_api_key,
             data=ChatCompletionRequest(
                 model=llm_config.model,
                 messages=[cast_message_to_subtype(m.to_openai_dict()) for m in messages],
@@ -455,7 +459,7 @@ def create(
             chat_completion_request=ChatCompletionRequest(
                 model="command-r-plus",  # TODO
                 messages=[cast_message_to_subtype(m.to_openai_dict()) for m in messages],
-                tools=[{"type": "function", "function": f} for f in functions] if functions else None,
+                tools=tools,
                 tool_choice=function_call,
                 # user=str(user_id),
                 # NOTE: max_tokens is required for Anthropic API
@@ -463,6 +467,60 @@ def create(
             ),
         )
 
+    elif llm_config.model_endpoint_type == "groq":
+        if stream:
+            raise NotImplementedError(f"Streaming not yet implemented for Groq.")
+
+        if model_settings.groq_api_key is None and llm_config.model_endpoint == "https://api.groq.com/openai/v1/chat/completions":
+            # only is a problem if we are *not* using an openai proxy
+            raise ValueError(f"Groq key is missing from letta config file")
+
+        # force to true for groq, since they don't support 'content' is non-null
+        inner_thoughts_in_kwargs = True
+        if inner_thoughts_in_kwargs:
+            functions = add_inner_thoughts_to_functions(
+                functions=functions,
+                inner_thoughts_key=INNER_THOUGHTS_KWARG,
+                inner_thoughts_description=INNER_THOUGHTS_KWARG_DESCRIPTION,
+            )
+
+        tools = [{"type": "function", "function": f} for f in functions] if functions is not None else None
+        data = ChatCompletionRequest(
+            model=llm_config.model,
+            messages=[m.to_openai_dict(put_inner_thoughts_in_kwargs=inner_thoughts_in_kwargs) for m in messages],
+            tools=tools,
+            tool_choice=function_call,
+            user=str(user_id),
+        )
+
+        # https://console.groq.com/docs/openai
+        # "The following fields are currently not supported and will result in a 400 error (yikes) if they are supplied:"
+        assert data.top_logprobs is None
+        assert data.logit_bias is None
+        assert data.logprobs == False
+        assert data.n == 1
+        # They mention that none of the messages can have names, but it seems to not error out (for now)
+
+        data.stream = False
+        if isinstance(stream_inferface, AgentChunkStreamingInterface):
+            stream_inferface.stream_start()
+        try:
+            # groq uses the openai chat completions API, so this component should be reusable
+            assert model_settings.groq_api_key is not None, "Groq key is missing"
+            response = openai_chat_completions_request(
+                url=llm_config.model_endpoint,
+                api_key=model_settings.groq_api_key,
+                chat_completion_request=data,
+            )
+        finally:
+            if isinstance(stream_inferface, AgentChunkStreamingInterface):
+                stream_inferface.stream_end()
+
+        if inner_thoughts_in_kwargs:
+            response = unpack_inner_thoughts_from_kwargs(response=response, inner_thoughts_key=INNER_THOUGHTS_KWARG)
+
+        return response
+
     # local model
     else:
         if stream:
@@ -481,6 +539,6 @@ def create(
             # hint
             first_message=first_message,
             # auth-related
-            auth_type=credentials.openllm_auth_type,
-            auth_key=credentials.openllm_key,
+            auth_type=model_settings.openllm_auth_type,
+            auth_key=model_settings.openllm_api_key,
         )

letta/local_llm/chat_completion_proxy.py

@@ -12,7 +12,6 @@ from letta.local_llm.grammars.gbnf_grammar_generator import (
     create_dynamic_model_from_function,
     generate_gbnf_grammar_and_documentation,
 )
-from letta.local_llm.groq.api import get_groq_completion
 from letta.local_llm.koboldcpp.api import get_koboldcpp_completion
 from letta.local_llm.llamacpp.api import get_llamacpp_completion
 from letta.local_llm.llm_chat_completion_wrappers import simple_summary_wrapper
@@ -170,8 +169,6 @@ def get_chat_completion(
             result, usage = get_ollama_completion(endpoint, auth_type, auth_key, model, prompt, context_window)
         elif endpoint_type == "vllm":
             result, usage = get_vllm_completion(endpoint, auth_type, auth_key, model, prompt, context_window, user)
-        elif endpoint_type == "groq":
-            result, usage = get_groq_completion(endpoint, auth_type, auth_key, model, prompt, context_window)
         else:
             raise LocalLLMError(
                 f"Invalid endpoint type {endpoint_type}, please set variable depending on your backend (webui, lmstudio, llamacpp, koboldcpp)"

letta/main.py

@@ -14,7 +14,7 @@ import letta.system as system
 # import benchmark
 from letta import create_client
 from letta.benchmark.benchmark import bench
-from letta.cli.cli import delete_agent, open_folder, quickstart, run, server, version
+from letta.cli.cli import delete_agent, open_folder, run, server, version
 from letta.cli.cli_config import add, add_tool, configure, delete, list, list_tools
 from letta.cli.cli_load import app as load_app
 from letta.config import LettaConfig
@@ -38,7 +38,6 @@ app.command(name="list-tools")(list_tools)
 app.command(name="delete")(delete)
 app.command(name="server")(server)
 app.command(name="folder")(open_folder)
-app.command(name="quickstart")(quickstart)
 # load data commands
 app.add_typer(load_app, name="load")
 # benchmark command

letta/metadata.py

@@ -151,6 +151,18 @@ class OrganizationModel(Base):
         return Organization(id=self.id, name=self.name, created_at=self.created_at)
 
 
+# TODO: eventually store providers?
+# class Provider(Base):
+#    __tablename__ = "providers"
+#    __table_args__ = {"extend_existing": True}
+#
+#    id = Column(String, primary_key=True)
+#    name = Column(String, nullable=False)
+#    created_at = Column(DateTime(timezone=True))
+#    api_key = Column(String, nullable=False)
+#    base_url = Column(String, nullable=False)
+
+
 class APIKeyModel(Base):
     """Data model for authentication tokens. One-to-many relationship with UserModel (1 User - N tokens)."""
 

letta/providers.py

@@ -0,0 +1,232 @@
+from typing import List, Optional
+
+from pydantic import BaseModel, Field
+
+from letta.constants import LLM_MAX_TOKENS
+from letta.schemas.embedding_config import EmbeddingConfig
+from letta.schemas.llm_config import LLMConfig
+
+
+class Provider(BaseModel):
+    base_url: str
+
+    def list_llm_models(self):
+        return []
+
+    def list_embedding_models(self):
+        return []
+
+    def get_model_context_window(self, model_name: str):
+        pass
+
+
+class OpenAIProvider(Provider):
+    name: str = "openai"
+    api_key: str = Field(..., description="API key for the OpenAI API.")
+    base_url: str = "https://api.openai.com/v1"
+
+    def list_llm_models(self) -> List[LLMConfig]:
+        from letta.llm_api.openai import openai_get_model_list
+
+        response = openai_get_model_list(self.base_url, api_key=self.api_key)
+        model_options = [obj["id"] for obj in response["data"]]
+
+        configs = []
+        for model_name in model_options:
+            context_window_size = self.get_model_context_window_size(model_name)
+
+            if not context_window_size:
+                continue
+            configs.append(
+                LLMConfig(model=model_name, model_endpoint_type="openai", model_endpoint=self.base_url, context_window=context_window_size)
+            )
+        return configs
+
+    def list_embedding_models(self) -> List[EmbeddingConfig]:
+
+        # TODO: actually automatically list models
+        return [
+            EmbeddingConfig(
+                embedding_model="text-embedding-ada-002",
+                embedding_endpoint_type="openai",
+                embedding_endpoint="https://api.openai.com/v1",
+                embedding_dim=1536,
+                embedding_chunk_size=300,
+            )
+        ]
+
+    def get_model_context_window_size(self, model_name: str):
+        if model_name in LLM_MAX_TOKENS:
+            return LLM_MAX_TOKENS[model_name]
+        else:
+            return None
+
+
+class AnthropicProvider(Provider):
+    name: str = "anthropic"
+    api_key: str = Field(..., description="API key for the Anthropic API.")
+    base_url: str = "https://api.anthropic.com/v1"
+
+    def list_llm_models(self) -> List[LLMConfig]:
+        from letta.llm_api.anthropic import anthropic_get_model_list
+
+        models = anthropic_get_model_list(self.base_url, api_key=self.api_key)
+
+        configs = []
+        for model in models:
+            configs.append(
+                LLMConfig(
+                    model=model["name"],
+                    model_endpoint_type="anthropic",
+                    model_endpoint=self.base_url,
+                    context_window=model["context_window"],
+                )
+            )
+        return configs
+
+    def list_embedding_models(self) -> List[EmbeddingConfig]:
+        return []
+
+
+class OllamaProvider(OpenAIProvider):
+    name: str = "ollama"
+    base_url: str = Field(..., description="Base URL for the Ollama API.")
+    api_key: Optional[str] = Field(None, description="API key for the Ollama API (default: `None`).")
+
+    def list_llm_models(self) -> List[LLMConfig]:
+        # https://github.com/ollama/ollama/blob/main/docs/api.md#list-local-models
+        import requests
+
+        response = requests.get(f"{self.base_url}/api/tags")
+        if response.status_code != 200:
+            raise Exception(f"Failed to list Ollama models: {response.text}")
+        response_json = response.json()
+
+        configs = []
+        for model in response_json["models"]:
+            context_window = self.get_model_context_window(model["name"])
+            configs.append(
+                LLMConfig(
+                    model=model["name"],
+                    model_endpoint_type="ollama",
+                    model_endpoint=self.base_url,
+                    context_window=context_window,
+                )
+            )
+        return configs
+
+    def get_model_context_window(self, model_name: str):
+
+        import requests
+
+        response = requests.post(f"{self.base_url}/api/show", json={"name": model_name, "verbose": True})
+        response_json = response.json()
+
+        # thank you vLLM: https://github.com/vllm-project/vllm/blob/main/vllm/config.py#L1675
+        possible_keys = [
+            # OPT
+            "max_position_embeddings",
+            # GPT-2
+            "n_positions",
+            # MPT
+            "max_seq_len",
+            # ChatGLM2
+            "seq_length",
+            # Command-R
+            "model_max_length",
+            # Others
+            "max_sequence_length",
+            "max_seq_length",
+            "seq_len",
+        ]
+
+        # max_position_embeddings
+        # parse model cards: nous, dolphon, llama
+        for key, value in response_json["model_info"].items():
+            if "context_window" in key:
+                return value
+        return None
+
+    def list_embedding_models(self) -> List[EmbeddingConfig]:
+        # TODO: filter embedding models
+        return []
+
+
+class GroqProvider(OpenAIProvider):
+    name: str = "groq"
+    base_url: str = "https://api.groq.com/openai/v1"
+    api_key: str = Field(..., description="API key for the Groq API.")
+
+    def list_llm_models(self) -> List[LLMConfig]:
+        from letta.llm_api.openai import openai_get_model_list
+
+        response = openai_get_model_list(self.base_url, api_key=self.api_key)
+        configs = []
+        for model in response["data"]:
+            if not "context_window" in model:
+                continue
+            configs.append(
+                LLMConfig(
+                    model=model["id"], model_endpoint_type="openai", model_endpoint=self.base_url, context_window=model["context_window"]
+                )
+            )
+        return configs
+
+    def list_embedding_models(self) -> List[EmbeddingConfig]:
+        return []
+
+    def get_model_context_window_size(self, model_name: str):
+        raise NotImplementedError
+
+
+class GoogleAIProvider(Provider):
+    # gemini
+    api_key: str = Field(..., description="API key for the Google AI API.")
+    service_endpoint: str = "generativelanguage"
+    base_url: str = "https://generativelanguage.googleapis.com"
+
+    def list_llm_models(self):
+        from letta.llm_api.google_ai import google_ai_get_model_list
+
+        # TODO: use base_url instead
+        model_options = google_ai_get_model_list(service_endpoint=self.service_endpoint, api_key=self.api_key)
+        model_options = [str(m["name"]) for m in model_options]
+        model_options = [mo[len("models/") :] if mo.startswith("models/") else mo for mo in model_options]
+        # TODO remove manual filtering for gemini-pro
+        model_options = [mo for mo in model_options if str(mo).startswith("gemini") and "-pro" in str(mo)]
+        # TODO: add context windows
+        # model_options = ["gemini-pro"]
+
+        configs = []
+        for model in model_options:
+            configs.append(
+                LLMConfig(
+                    model=model,
+                    model_endpoint_type="google_ai",
+                    model_endpoint=self.base_url,
+                    context_window=self.get_model_context_window(model),
+                )
+            )
+        return configs
+
+    def list_embedding_models(self):
+        return []
+
+    def get_model_context_window(self, model_name: str):
+        from letta.llm_api.google_ai import google_ai_get_model_context_window
+
+        # TODO: use base_url instead
+        return google_ai_get_model_context_window(self.service_endpoint, self.api_key, model_name)
+
+
+class AzureProvider(Provider):
+    pass
+
+
+class VLLMProvider(OpenAIProvider):
+    # NOTE: vLLM only serves one model at a time (so could configure that through env variables)
+    pass
+
+
+class CohereProvider(OpenAIProvider):
+    pass

letta/schemas/block.py

@@ -53,7 +53,7 @@ class BaseBlock(LettaBase, validate_assignment=True):
         super().__setattr__(name, value)
         if name == "value":
             # run validation
-            self.__class__.validate(self.dict(exclude_unset=True))
+            self.__class__.model_validate(self.model_dump(exclude_unset=True))
 
 
 class Block(BaseBlock):

letta/schemas/letta_request.py

@@ -2,6 +2,7 @@ from typing import List
 
 from pydantic import BaseModel, Field
 
+from letta.constants import DEFAULT_MESSAGE_TOOL, DEFAULT_MESSAGE_TOOL_KWARG
 from letta.schemas.message import MessageCreate
 
 
@@ -21,3 +22,19 @@ class LettaRequest(BaseModel):
         default=False,
         description="Set True to return the raw Message object. Set False to return the Message in the format of the Letta API.",
     )
+
+    # Flags to support the use of AssistantMessage message types
+
+    use_assistant_message: bool = Field(
+        default=False,
+        description="[Only applicable if return_message_object is False] If true, returns AssistantMessage objects when the agent calls a designated message tool. If false, return FunctionCallMessage objects for all tool calls.",
+    )
+
+    assistant_message_function_name: str = Field(
+        default=DEFAULT_MESSAGE_TOOL,
+        description="[Only applicable if use_assistant_message is True] The name of the designated message tool.",
+    )
+    assistant_message_function_kwarg: str = Field(
+        default=DEFAULT_MESSAGE_TOOL_KWARG,
+        description="[Only applicable if use_assistant_message is True] The name of the message argument in the designated message tool.",
+    )

letta/schemas/letta_response.py

@@ -6,6 +6,7 @@ from letta.schemas.enums import MessageStreamStatus
 from letta.schemas.letta_message import LettaMessage
 from letta.schemas.message import Message
 from letta.schemas.usage import LettaUsageStatistics
+from letta.utils import json_dumps
 
 # TODO: consider moving into own file
 
@@ -23,6 +24,16 @@ class LettaResponse(BaseModel):
     messages: Union[List[Message], List[LettaMessage]] = Field(..., description="The messages returned by the agent.")
     usage: LettaUsageStatistics = Field(..., description="The usage statistics of the agent.")
 
+    def __str__(self):
+        return json_dumps(
+            {
+                "messages": [message.model_dump() for message in self.messages],
+                # Assume `Message` and `LettaMessage` have a `dict()` method
+                "usage": self.usage.model_dump(),  # Assume `LettaUsageStatistics` has a `dict()` method
+            },
+            indent=4,
+        )
+
 
 # The streaming response is either [DONE], [DONE_STEP], [DONE], an error, or a LettaMessage
 LettaStreamingResponse = Union[LettaMessage, MessageStreamStatus]

letta/schemas/llm_config.py

@@ -1,4 +1,4 @@
-from typing import Optional
+from typing import Literal, Optional
 
 from pydantic import BaseModel, ConfigDict, Field
 
@@ -17,7 +17,23 @@ class LLMConfig(BaseModel):
 
     # TODO: 🤮 don't default to a vendor! bug city!
     model: str = Field(..., description="LLM model name. ")
-    model_endpoint_type: str = Field(..., description="The endpoint type for the model.")
+    model_endpoint_type: Literal[
+        "openai",
+        "anthropic",
+        "cohere",
+        "google_ai",
+        "azure",
+        "groq",
+        "ollama",
+        "webui",
+        "webui-legacy",
+        "lmstudio",
+        "lmstudio-legacy",
+        "llamacpp",
+        "koboldcpp",
+        "vllm",
+        "hugging-face",
+    ] = Field(..., description="The endpoint type for the model.")
     model_endpoint: str = Field(..., description="The endpoint for the model.")
     model_wrapper: Optional[str] = Field(None, description="The wrapper for the model.")
     context_window: int = Field(..., description="The context window size for the model.")