letta-nightly 0.6.4.dev20241217104233__tar.gz → 0.6.5.dev20241218055539__tar.gz

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: letta-nightly
-Version: 0.6.4.dev20241217104233
+Version: 0.6.5.dev20241218055539
 Summary: Create LLM agents with long-term memory and custom tools
 License: Apache License
 Author: Letta Team

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.6.4"
+__version__ = "0.6.5"
 
 # import clients
 from letta.client.client import LocalClient, RESTClient, create_client

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/agent.py

@@ -18,6 +18,7 @@ from letta.constants import (
     MESSAGE_SUMMARY_WARNING_FRAC,
     O1_BASE_TOOLS,
     REQ_HEARTBEAT_MESSAGE,
+    STRUCTURED_OUTPUT_MODELS,
 )
 from letta.errors import LLMError
 from letta.helpers import ToolRulesSolver
@@ -63,6 +64,7 @@ from letta.system import (
 )
 from letta.utils import (
     count_tokens,
+    get_friendly_error_msg,
     get_local_time,
     get_tool_call_id,
     get_utc_time,
@@ -258,9 +260,6 @@ class Agent(BaseAgent):
 
         self.user = user
 
-        # link tools
-        self.link_tools(agent_state.tools)
-
         # initialize a tool rules solver
         if agent_state.tool_rules:
             # if there are tool rules, print out a warning
@@ -276,6 +275,7 @@ class Agent(BaseAgent):
 
         # gpt-4, gpt-3.5-turbo, ...
         self.model = self.agent_state.llm_config.model
+        self.check_tool_rules()
 
         # state managers
         self.block_manager = BlockManager()
@@ -295,8 +295,6 @@ class Agent(BaseAgent):
         self.agent_manager = AgentManager()
 
         # State needed for heartbeat pausing
-        self.pause_heartbeats_start = None
-        self.pause_heartbeats_minutes = 0
 
         self.first_message_verify_mono = first_message_verify_mono
 
@@ -381,6 +379,16 @@ class Agent(BaseAgent):
         # Create the agent in the DB
         self.update_state()
 
+    def check_tool_rules(self):
+        if self.model not in STRUCTURED_OUTPUT_MODELS:
+            if len(self.tool_rules_solver.init_tool_rules) > 1:
+                raise ValueError(
+                    "Multiple initial tools are not supported for non-structured models. Please use only one initial tool rule."
+                )
+            self.supports_structured_output = False
+        else:
+            self.supports_structured_output = True
+
     def update_memory_if_change(self, new_memory: Memory) -> bool:
         """
         Update internal memory object and system prompt if there have been modifications.
@@ -415,11 +423,21 @@ class Agent(BaseAgent):
             return True
         return False
 
-    def execute_tool_and_persist_state(self, function_name, function_to_call, function_args):
+    def execute_tool_and_persist_state(self, function_name: str, function_args: dict, target_letta_tool: Tool):
         """
         Execute tool modifications and persist the state of the agent.
         Note: only some agent state modifications will be persisted, such as data in the AgentState ORM and block data
         """
+        # TODO: Get rid of this. This whole piece is pretty shady, that we exec the function to just get the type hints for args.
+        env = {}
+        env.update(globals())
+        exec(target_letta_tool.source_code, env)
+        callable_func = env[target_letta_tool.json_schema["name"]]
+        spec = inspect.getfullargspec(callable_func).annotations
+        for name, arg in function_args.items():
+            if isinstance(function_args[name], dict):
+                function_args[name] = spec[name](**function_args[name])
+
         # TODO: add agent manager here
         orig_memory_str = self.agent_state.memory.compile()
 
@@ -432,11 +450,11 @@ class Agent(BaseAgent):
             if function_name in BASE_TOOLS or function_name in O1_BASE_TOOLS:
                 # base tools are allowed to access the `Agent` object and run on the database
                 function_args["self"] = self  # need to attach self to arg since it's dynamically linked
-                function_response = function_to_call(**function_args)
+                function_response = callable_func(**function_args)
             else:
                 # execute tool in a sandbox
                 # TODO: allow agent_state to specify which sandbox to execute tools in
-                sandbox_run_result = ToolExecutionSandbox(function_name, function_args, self.agent_state.created_by_id).run(
+                sandbox_run_result = ToolExecutionSandbox(function_name, function_args, self.user).run(
                     agent_state=self.agent_state.__deepcopy__()
                 )
                 function_response, updated_agent_state = sandbox_run_result.func_return, sandbox_run_result.agent_state
@@ -446,12 +464,9 @@ class Agent(BaseAgent):
         except Exception as e:
             # Need to catch error here, or else trunction wont happen
             # TODO: modify to function execution error
-            from letta.constants import MAX_ERROR_MESSAGE_CHAR_LIMIT
-
-            error_msg = f"Error executing tool {function_name}: {e}"
-            if len(error_msg) > MAX_ERROR_MESSAGE_CHAR_LIMIT:
-                error_msg = error_msg[:MAX_ERROR_MESSAGE_CHAR_LIMIT]
-            raise ValueError(error_msg)
+            function_response = get_friendly_error_msg(
+                function_name=function_name, exception_name=type(e).__name__, exception_message=str(e)
+            )
 
         return function_response
 
@@ -464,27 +479,6 @@ class Agent(BaseAgent):
     def messages(self, value):
         raise Exception("Modifying message list directly not allowed")
 
-    def link_tools(self, tools: List[Tool]):
-        """Bind a tool object (schema + python function) to the agent object"""
-
-        # Store the functions schemas (this is passed as an argument to ChatCompletion)
-        self.functions = []
-        self.functions_python = {}
-        env = {}
-        env.update(globals())
-        for tool in tools:
-            try:
-                # WARNING: name may not be consistent?
-                # if tool.module:  # execute the whole module
-                #    exec(tool.module, env)
-                # else:
-                exec(tool.source_code, env)
-                self.functions_python[tool.json_schema["name"]] = env[tool.json_schema["name"]]
-                self.functions.append(tool.json_schema)
-            except Exception:
-                warnings.warn(f"WARNING: tool {tool.name} failed to link")
-        assert all([callable(f) for k, f in self.functions_python.items()]), self.functions_python
-
     def _load_messages_from_recall(self, message_ids: List[str]) -> List[Message]:
         """Load a list of messages from recall storage"""
 
@@ -588,14 +582,32 @@ class Agent(BaseAgent):
         empty_response_retry_limit: int = 3,
         backoff_factor: float = 0.5,  # delay multiplier for exponential backoff
         max_delay: float = 10.0,  # max delay between retries
+        step_count: Optional[int] = None,
     ) -> ChatCompletionResponse:
         """Get response from LLM API with robust retry mechanism."""
 
         allowed_tool_names = self.tool_rules_solver.get_allowed_tool_names()
+        agent_state_tool_jsons = [t.json_schema for t in self.agent_state.tools]
+
         allowed_functions = (
-            self.functions if not allowed_tool_names else [func for func in self.functions if func["name"] in allowed_tool_names]
+            agent_state_tool_jsons
+            if not allowed_tool_names
+            else [func for func in agent_state_tool_jsons if func["name"] in allowed_tool_names]
         )
 
+        # For the first message, force the initial tool if one is specified
+        force_tool_call = None
+        if (
+            step_count is not None
+            and step_count == 0
+            and not self.supports_structured_output
+            and len(self.tool_rules_solver.init_tool_rules) > 0
+        ):
+            force_tool_call = self.tool_rules_solver.init_tool_rules[0].tool_name
+        # Force a tool call if exactly one tool is specified
+        elif step_count is not None and step_count > 0 and len(allowed_tool_names) == 1:
+            force_tool_call = allowed_tool_names[0]
+
         for attempt in range(1, empty_response_retry_limit + 1):
             try:
                 response = create(
@@ -603,9 +615,10 @@ class Agent(BaseAgent):
                     messages=message_sequence,
                     user_id=self.agent_state.created_by_id,
                     functions=allowed_functions,
-                    functions_python=self.functions_python,
+                    # functions_python=self.functions_python, do we need this?
                     function_call=function_call,
                     first_message=first_message,
+                    force_tool_call=force_tool_call,
                     stream=stream,
                     stream_interface=self.interface,
                 )
@@ -711,10 +724,13 @@ class Agent(BaseAgent):
             function_name = function_call.name
             printd(f"Request to call function {function_name} with tool_call_id: {tool_call_id}")
 
-            # Failure case 1: function name is wrong
-            try:
-                function_to_call = self.functions_python[function_name]
-            except KeyError:
+            # Failure case 1: function name is wrong (not in agent_state.tools)
+            target_letta_tool = None
+            for t in self.agent_state.tools:
+                if t.name == function_name:
+                    target_letta_tool = t
+
+            if not target_letta_tool:
                 error_msg = f"No function named {function_name}"
                 function_response = package_function_response(False, error_msg)
                 messages.append(
@@ -782,14 +798,8 @@ class Agent(BaseAgent):
             #       this is because the function/tool role message is only created once the function/tool has executed/returned
             self.interface.function_message(f"Running {function_name}({function_args})", msg_obj=messages[-1])
             try:
-                spec = inspect.getfullargspec(function_to_call).annotations
-
-                for name, arg in function_args.items():
-                    if isinstance(function_args[name], dict):
-                        function_args[name] = spec[name](**function_args[name])
-
                 # handle tool execution (sandbox) and state updates
-                function_response = self.execute_tool_and_persist_state(function_name, function_to_call, function_args)
+                function_response = self.execute_tool_and_persist_state(function_name, function_args, target_letta_tool)
 
                 # handle trunction
                 if function_name in ["conversation_search", "conversation_search_date", "archival_memory_search"]:
@@ -801,8 +811,7 @@ class Agent(BaseAgent):
                     truncate = True
 
                 # get the function response limit
-                tool_obj = [tool for tool in self.agent_state.tools if tool.name == function_name][0]
-                return_char_limit = tool_obj.return_char_limit
+                return_char_limit = target_letta_tool.return_char_limit
                 function_response_string = validate_function_response(
                     function_response, return_char_limit=return_char_limit, truncate=truncate
                 )
@@ -897,6 +906,7 @@ class Agent(BaseAgent):
         step_count = 0
         while True:
             kwargs["first_message"] = False
+            kwargs["step_count"] = step_count
             step_response = self.inner_step(
                 messages=next_input_message,
                 **kwargs,
@@ -972,6 +982,7 @@ class Agent(BaseAgent):
         first_message_retry_limit: int = FIRST_MESSAGE_ATTEMPTS,
         skip_verify: bool = False,
         stream: bool = False,  # TODO move to config?
+        step_count: Optional[int] = None,
     ) -> AgentStepResponse:
         """Runs a single step in the agent loop (generates at most one LLM call)"""
 
@@ -1014,7 +1025,9 @@ class Agent(BaseAgent):
             else:
                 response = self._get_ai_reply(
                     message_sequence=input_message_sequence,
+                    first_message=first_message,
                     stream=stream,
+                    step_count=step_count,
                 )
 
             # Step 3: check if LLM wanted to call a function
@@ -1235,17 +1248,6 @@ class Agent(BaseAgent):
 
         printd(f"Ran summarizer, messages length {prior_len} -> {len(self.messages)}")
 
-    def heartbeat_is_paused(self):
-        """Check if there's a requested pause on timed heartbeats"""
-
-        # Check if the pause has been initiated
-        if self.pause_heartbeats_start is None:
-            return False
-
-        # Check if it's been more than pause_heartbeats_minutes since pause_heartbeats_start
-        elapsed_time = get_utc_time() - self.pause_heartbeats_start
-        return elapsed_time.total_seconds() < self.pause_heartbeats_minutes * 60
-
     def _swap_system_message_in_buffer(self, new_system_message: str):
         """Update the system message (NOT prompt) of the Agent (requires updating the internal buffer)"""
         assert isinstance(new_system_message, str)
@@ -1370,7 +1372,7 @@ class Agent(BaseAgent):
         agent_manager: AgentManager,
     ):
         """Attach a source to the agent using the SourcesAgents ORM relationship.
- 
+
         Args:
             user: User performing the action
             source_id: ID of the source to attach
@@ -1553,9 +1555,10 @@ class Agent(BaseAgent):
         num_tokens_external_memory_summary = count_tokens(external_memory_summary)
 
         # tokens taken up by function definitions
-        if self.functions:
-            available_functions_definitions = [ChatCompletionRequestTool(type="function", function=f) for f in self.functions]
-            num_tokens_available_functions_definitions = num_tokens_from_functions(functions=self.functions, model=self.model)
+        agent_state_tool_jsons = [t.json_schema for t in self.agent_state.tools]
+        if agent_state_tool_jsons:
+            available_functions_definitions = [ChatCompletionRequestTool(type="function", function=f) for f in agent_state_tool_jsons]
+            num_tokens_available_functions_definitions = num_tokens_from_functions(functions=agent_state_tool_jsons, model=self.model)
         else:
             available_functions_definitions = []
             num_tokens_available_functions_definitions = 0

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/client/client.py

@@ -2156,6 +2156,7 @@ class LocalClient(AbstractClient):
             "block_ids": [b.id for b in memory.get_blocks()] + block_ids,
             "tool_ids": tool_ids,
             "tool_rules": tool_rules,
+            "include_base_tools": include_base_tools,
             "system": system,
             "agent_type": agent_type,
             "llm_config": llm_config if llm_config else self._default_llm_config,

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/constants.py

@@ -23,6 +23,7 @@ MIN_CONTEXT_WINDOW = 4096
 
 # embeddings
 MAX_EMBEDDING_DIM = 4096  # maximum supported embeding size - do NOT change or else DBs will need to be reset
+DEFAULT_EMBEDDING_CHUNK_SIZE = 300
 
 # tokenizers
 EMBEDDING_TO_TOKENIZER_MAP = {
@@ -37,7 +38,8 @@ DEFAULT_HUMAN = "basic"
 DEFAULT_PRESET = "memgpt_chat"
 
 # Base tools that cannot be edited, as they access agent state directly
-BASE_TOOLS = ["send_message", "conversation_search", "conversation_search_date", "archival_memory_insert", "archival_memory_search"]
+# Note that we don't include "conversation_search_date" for now
+BASE_TOOLS = ["send_message", "conversation_search", "archival_memory_insert", "archival_memory_search"]
 O1_BASE_TOOLS = ["send_thinking_message", "send_final_message"]
 # Base memory tools CAN be edited, and are added by default by the server
 BASE_MEMORY_TOOLS = ["core_memory_append", "core_memory_replace"]
@@ -48,6 +50,9 @@ BASE_MEMORY_TOOLS = ["core_memory_append", "core_memory_replace"]
 DEFAULT_MESSAGE_TOOL = "send_message"
 DEFAULT_MESSAGE_TOOL_KWARG = "message"
 
+# Structured output models
+STRUCTURED_OUTPUT_MODELS = {"gpt-4o", "gpt-4o-mini"}
+
 # LOGGER_LOG_LEVEL is use to convert Text to Logging level value for logging mostly for Cli input to setting level
 LOGGER_LOG_LEVELS = {"CRITICAL": CRITICAL, "ERROR": ERROR, "WARN": WARN, "WARNING": WARNING, "INFO": INFO, "DEBUG": DEBUG, "NOTSET": NOTSET}
 

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/embeddings.py

@@ -234,16 +234,10 @@ def embedding_model(config: EmbeddingConfig, user_id: Optional[uuid.UUID] = None
         )
     elif endpoint_type == "ollama":
 
-        from llama_index.embeddings.ollama import OllamaEmbedding
-
-        ollama_additional_kwargs = {}
-        callback_manager = None
-
-        model = OllamaEmbedding(
-            model_name=config.embedding_model,
+        model = OllamaEmbeddings(
+            model=config.embedding_model,
             base_url=config.embedding_endpoint,
-            ollama_additional_kwargs=ollama_additional_kwargs or {},
-            callback_manager=callback_manager or None,
+            ollama_additional_kwargs={},
         )
         return model
 

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/functions/function_sets/base.py

@@ -1,16 +1,6 @@
-from datetime import datetime
 from typing import Optional
 
 from letta.agent import Agent
-from letta.constants import MAX_PAUSE_HEARTBEATS
-from letta.services.agent_manager import AgentManager
-
-# import math
-# from letta.utils import json_dumps
-
-### Functions / tools the agent can use
-# All functions should return a response string (or None)
-# If the function fails, throw an exception
 
 
 def send_message(self: "Agent", message: str) -> Optional[str]:
@@ -28,36 +18,6 @@ def send_message(self: "Agent", message: str) -> Optional[str]:
     return None
 
 
-# Construct the docstring dynamically (since it should use the external constants)
-pause_heartbeats_docstring = f"""
-Temporarily ignore timed heartbeats. You may still receive messages from manual heartbeats and other events.
-
-Args:
-    minutes (int): Number of minutes to ignore heartbeats for. Max value of {MAX_PAUSE_HEARTBEATS} minutes ({MAX_PAUSE_HEARTBEATS // 60} hours).
-
-Returns:
-    str: Function status response
-"""
-
-
-def pause_heartbeats(self: "Agent", minutes: int) -> Optional[str]:
-    import datetime
-
-    from letta.constants import MAX_PAUSE_HEARTBEATS
-
-    minutes = min(MAX_PAUSE_HEARTBEATS, minutes)
-
-    # Record the current time
-    self.pause_heartbeats_start = datetime.datetime.now(datetime.timezone.utc)
-    # And record how long the pause should go for
-    self.pause_heartbeats_minutes = int(minutes)
-
-    return f"Pausing timed heartbeats for {minutes} min"
-
-
-pause_heartbeats.__doc__ = pause_heartbeats_docstring
-
-
 def conversation_search(self: "Agent", query: str, page: Optional[int] = 0) -> Optional[str]:
     """
     Search prior conversation history using case-insensitive string matching.
@@ -84,19 +44,19 @@ def conversation_search(self: "Agent", query: str, page: Optional[int] = 0) -> O
     count = RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE
     # TODO: add paging by page number. currently cursor only works with strings.
     # original: start=page * count
-    results = self.message_manager.list_user_messages_for_agent(
+    messages = self.message_manager.list_user_messages_for_agent(
         agent_id=self.agent_state.id,
         actor=self.user,
         query_text=query,
         limit=count,
     )
-    total = len(results)
+    total = len(messages)
     num_pages = math.ceil(total / count) - 1  # 0 index
-    if len(results) == 0:
+    if len(messages) == 0:
         results_str = f"No results found."
     else:
-        results_pref = f"Showing {len(results)} of {total} results (page {page}/{num_pages}):"
-        results_formatted = [f"timestamp: {d['timestamp']}, {d['message']['role']} - {d['message']['content']}" for d in results]
+        results_pref = f"Showing {len(messages)} of {total} results (page {page}/{num_pages}):"
+        results_formatted = [message.text for message in messages]
         results_str = f"{results_pref} {json_dumps(results_formatted)}"
     return results_str
 
@@ -114,6 +74,7 @@ def conversation_search_date(self: "Agent", start_date: str, end_date: str, page
         str: Query result string
     """
     import math
+    from datetime import datetime
 
     from letta.constants import RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE
     from letta.utils import json_dumps
@@ -142,7 +103,6 @@ def conversation_search_date(self: "Agent", start_date: str, end_date: str, page
         start_date=start_datetime,
         end_date=end_datetime,
         limit=count,
-        # start_date=start_date, end_date=end_date, limit=count, start=page * count
     )
     total = len(results)
     num_pages = math.ceil(total / count) - 1  # 0 index
@@ -186,10 +146,8 @@ def archival_memory_search(self: "Agent", query: str, page: Optional[int] = 0, s
     Returns:
         str: Query result string
     """
-    import math
 
     from letta.constants import RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE
-    from letta.utils import json_dumps
 
     if page is None or (isinstance(page, str) and page.lower().strip() == "none"):
         page = 0
@@ -198,7 +156,7 @@ def archival_memory_search(self: "Agent", query: str, page: Optional[int] = 0, s
     except:
         raise ValueError(f"'page' argument must be an integer")
     count = RETRIEVAL_QUERY_DEFAULT_PAGE_SIZE
-    
+
     try:
         # Get results using passage manager
         all_results = self.agent_manager.list_passages(
@@ -207,7 +165,7 @@ def archival_memory_search(self: "Agent", query: str, page: Optional[int] = 0, s
             query_text=query,
             limit=count + start,  # Request enough results to handle offset
             embedding_config=self.agent_state.embedding_config,
-            embed_query=True
+            embed_query=True,
         )
 
         # Apply pagination
@@ -215,13 +173,7 @@ def archival_memory_search(self: "Agent", query: str, page: Optional[int] = 0, s
         paged_results = all_results[start:end]
 
         # Format results to match previous implementation
-        formatted_results = [
-            {
-                "timestamp": str(result.created_at),
-                "content": result.text
-            }
-            for result in paged_results
-        ]
+        formatted_results = [{"timestamp": str(result.created_at), "content": result.text} for result in paged_results]
 
         return formatted_results, len(formatted_results)
 

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/functions/schema_generator.py

@@ -386,7 +386,7 @@ def generate_schema(function, name: Optional[str] = None, description: Optional[
     # append the heartbeat
     # TODO: don't hard-code
     # TODO: if terminal, don't include this
-    if function.__name__ not in ["send_message", "pause_heartbeats"]:
+    if function.__name__ not in ["send_message"]:
         schema["parameters"]["properties"]["request_heartbeat"] = {
             "type": "boolean",
             "description": "Request an immediate heartbeat after function execution. Set to `True` if you want to send a follow-up message or run a follow-up function.",

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/llm_api/anthropic.py

@@ -99,16 +99,20 @@ def convert_tools_to_anthropic_format(tools: List[Tool]) -> List[dict]:
         - 1 level less of nesting
         - "parameters" -> "input_schema"
     """
-    tools_dict_list = []
+    formatted_tools = []
     for tool in tools:
-        tools_dict_list.append(
-            {
-                "name": tool.function.name,
-                "description": tool.function.description,
-                "input_schema": tool.function.parameters,
+        formatted_tool = {
+            "name"         : tool.function.name,
+            "description"  : tool.function.description,
+            "input_schema"   : tool.function.parameters or {
+                "type": "object",
+                "properties": {},
+                "required": []
             }
-        )
-    return tools_dict_list
+        }
+        formatted_tools.append(formatted_tool)
+
+    return formatted_tools
 
 
 def merge_tool_results_into_user_messages(messages: List[dict]):
@@ -258,10 +262,24 @@ def convert_anthropic_response_to_chatcompletion(
                     ),
                 )
             ]
-        else:
-            # Just inner mono
-            content = strip_xml_tags(string=response_json["content"][0]["text"], tag=inner_thoughts_xml_tag)
-            tool_calls = None
+        elif len(response_json["content"]) == 1:
+            if response_json["content"][0]["type"] == "tool_use":
+                # function call only
+                content = None
+                tool_calls = [
+                    ToolCall(
+                        id=response_json["content"][0]["id"],
+                        type="function",
+                        function=FunctionCall(
+                            name=response_json["content"][0]["name"],
+                            arguments=json.dumps(response_json["content"][0]["input"], indent=2),
+                        ),
+                    )
+                ]
+            else:
+                # inner mono only
+                content = strip_xml_tags(string=response_json["content"][0]["text"], tag=inner_thoughts_xml_tag)
+                tool_calls = None
     else:
         raise RuntimeError("Unexpected type for content in response_json.")
 
@@ -323,6 +341,14 @@ def anthropic_chat_completions_request(
     if anthropic_tools is not None:
         data["tools"] = anthropic_tools
 
+        # TODO: Add support for other tool_choice options like "auto", "any"
+        if len(anthropic_tools) == 1:
+            data["tool_choice"] = {
+                "type": "tool",  # Changed from "function" to "tool"
+                "name": anthropic_tools[0]["name"],  # Directly specify name without nested "function" object
+                "disable_parallel_tool_use": True  # Force single tool use
+            }
+
     # Move 'system' to the top level
     # 'messages: Unexpected role "system". The Messages API accepts a top-level `system` parameter, not "system" as an input message role.'
     assert data["messages"][0]["role"] == "system", f"Expected 'system' role in messages[0]:\n{data['messages'][0]}"
@@ -358,7 +384,6 @@ def anthropic_chat_completions_request(
     data.pop("top_p", None)
     data.pop("presence_penalty", None)
     data.pop("user", None)
-    data.pop("tool_choice", None)
 
     response_json = make_post_request(url, headers, data)
     return convert_anthropic_response_to_chatcompletion(response_json=response_json, inner_thoughts_xml_tag=inner_thoughts_xml_tag)

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/llm_api/llm_api_tools.py

@@ -113,6 +113,7 @@ def create(
     function_call: str = "auto",
     # hint
     first_message: bool = False,
+    force_tool_call: Optional[str] = None,  # Force a specific tool to be called
     # use tool naming?
     # if false, will use deprecated 'functions' style
     use_tool_naming: bool = True,
@@ -252,6 +253,16 @@ def create(
         if not use_tool_naming:
             raise NotImplementedError("Only tool calling supported on Anthropic API requests")
 
+        tool_call = None
+        if force_tool_call is not None:
+            tool_call = {
+                "type": "function",
+                "function": {
+                    "name": force_tool_call
+                }
+            }
+            assert functions is not None
+
         return anthropic_chat_completions_request(
             url=llm_config.model_endpoint,
             api_key=model_settings.anthropic_api_key,
@@ -259,7 +270,7 @@ def create(
                 model=llm_config.model,
                 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,
-                # tool_choice=function_call,
+                tool_choice=tool_call,
                 # user=str(user_id),
                 # NOTE: max_tokens is required for Anthropic API
                 max_tokens=1024,  # TODO make dynamic

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/local_llm/function_parser.py

@@ -3,7 +3,7 @@ import json
 
 from letta.utils import json_dumps, json_loads
 
-NO_HEARTBEAT_FUNCS = ["send_message", "pause_heartbeats"]
+NO_HEARTBEAT_FUNCS = ["send_message"]
 
 
 def insert_heartbeat(message):

{letta_nightly-0.6.4.dev20241217104233 → letta_nightly-0.6.5.dev20241218055539}/letta/orm/errors.py

@@ -12,3 +12,11 @@ class UniqueConstraintViolationError(ValueError):
 
 class ForeignKeyConstraintViolationError(ValueError):
     """Custom exception for foreign key constraint violations."""
+
+
+class DatabaseTimeoutError(Exception):
+    """Custom exception for database timeout issues."""
+
+    def __init__(self, message="Database operation timed out", original_exception=None):
+        super().__init__(message)
+        self.original_exception = original_exception