letta-nightly 0.6.27.dev20250220104103__py3-none-any.whl → 0.6.29.dev20250221033538__py3-none-any.whl

letta/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.6.27"
+__version__ = "0.6.29"
 
 # import clients
 from letta.client.client import LocalClient, RESTClient, create_client

letta/agent.py

@@ -60,6 +60,7 @@ from letta.services.tool_manager import ToolManager
 from letta.settings import summarizer_settings
 from letta.streaming_interface import StreamingRefreshCLIInterface
 from letta.system import get_heartbeat, get_token_limit_warning, package_function_response, package_summarize_message, package_user_message
+from letta.tracing import trace_method
 from letta.utils import (
     count_tokens,
     get_friendly_error_msg,
@@ -309,6 +310,7 @@ class Agent(BaseAgent):
         # Return updated messages
         return messages
 
+    @trace_method("Get AI Reply")
     def _get_ai_reply(
         self,
         message_sequence: List[Message],
@@ -320,6 +322,7 @@ class Agent(BaseAgent):
         max_delay: float = 10.0,  # max delay between retries
         step_count: Optional[int] = None,
         last_function_failed: bool = False,
+        put_inner_thoughts_first: bool = True,
     ) -> ChatCompletionResponse:
         """Get response from LLM API with robust retry mechanism."""
         log_telemetry(self.logger, "_get_ai_reply start")
@@ -365,6 +368,7 @@ class Agent(BaseAgent):
                     force_tool_call=force_tool_call,
                     stream=stream,
                     stream_interface=self.interface,
+                    put_inner_thoughts_first=put_inner_thoughts_first,
                 )
                 log_telemetry(self.logger, "_get_ai_reply create finish")
 
@@ -399,6 +403,7 @@ class Agent(BaseAgent):
         log_telemetry(self.logger, "_handle_ai_response finish catch-all exception")
         raise Exception("Retries exhausted and no valid response received.")
 
+    @trace_method("Handle AI Response")
     def _handle_ai_response(
         self,
         response_message: ChatCompletionMessage,  # TODO should we eventually move the Message creation outside of this function?
@@ -492,7 +497,10 @@ class Agent(BaseAgent):
             try:
                 raw_function_args = function_call.arguments
                 function_args = parse_json(raw_function_args)
-            except Exception:
+                if not isinstance(function_args, dict):
+                    raise ValueError(f"Function arguments are not a dictionary: {function_args} (raw={raw_function_args})")
+            except Exception as e:
+                print(e)
                 error_msg = f"Error parsing JSON for function '{function_name}' arguments: {function_call.arguments}"
                 function_response = "None"  # more like "never ran?"
                 messages = self._handle_function_error_response(
@@ -627,15 +635,22 @@ class Agent(BaseAgent):
         elif self.tool_rules_solver.is_terminal_tool(function_name):
             heartbeat_request = False
 
+        # if continue tool rule, then  must request a heartbeat
+        # TODO: dont even include heartbeats in the args
+        if self.tool_rules_solver.is_continue_tool(function_name):
+            heartbeat_request = True
+
         log_telemetry(self.logger, "_handle_ai_response finish")
         return messages, heartbeat_request, function_failed
 
+    @trace_method("Agent Step")
     def step(
         self,
         messages: Union[Message, List[Message]],
         # additional args
         chaining: bool = True,
         max_chaining_steps: Optional[int] = None,
+        put_inner_thoughts_first: bool = True,
         **kwargs,
     ) -> LettaUsageStatistics:
         """Run Agent.step in a loop, handling chaining via heartbeat requests and function failures"""
@@ -650,6 +665,7 @@ class Agent(BaseAgent):
             kwargs["last_function_failed"] = function_failed
             step_response = self.inner_step(
                 messages=next_input_message,
+                put_inner_thoughts_first=put_inner_thoughts_first,
                 **kwargs,
             )
 
@@ -731,9 +747,9 @@ class Agent(BaseAgent):
         metadata: Optional[dict] = None,
         summarize_attempt_count: int = 0,
         last_function_failed: bool = False,
+        put_inner_thoughts_first: bool = True,
     ) -> AgentStepResponse:
         """Runs a single step in the agent loop (generates at most one LLM call)"""
-
         try:
 
             # Extract job_id from metadata if present
@@ -766,6 +782,7 @@ class Agent(BaseAgent):
                 stream=stream,
                 step_count=step_count,
                 last_function_failed=last_function_failed,
+                put_inner_thoughts_first=put_inner_thoughts_first,
             )
             if not response:
                 # EDGE CASE: Function call failed AND there's no tools left for agent to call -> return early

letta/client/client.py

@@ -2351,6 +2351,7 @@ class LocalClient(AbstractClient):
         tool_rules: Optional[List[BaseToolRule]] = None,
         include_base_tools: Optional[bool] = True,
         include_multi_agent_tools: bool = False,
+        include_base_tool_rules: bool = True,
         # metadata
         metadata: Optional[Dict] = {"human:": DEFAULT_HUMAN, "persona": DEFAULT_PERSONA},
         description: Optional[str] = None,
@@ -2402,6 +2403,7 @@ class LocalClient(AbstractClient):
             "tool_rules": tool_rules,
             "include_base_tools": include_base_tools,
             "include_multi_agent_tools": include_multi_agent_tools,
+            "include_base_tool_rules": include_base_tool_rules,
             "system": system,
             "agent_type": agent_type,
             "llm_config": llm_config if llm_config else self._default_llm_config,

letta/constants.py

@@ -86,6 +86,8 @@ NON_USER_MSG_PREFIX = "[This is an automated system message hidden from the user
 # The max amount of tokens supported by the underlying model (eg 8k for gpt-4 and Mistral 7B)
 LLM_MAX_TOKENS = {
     "DEFAULT": 8192,
+    "deepseek-chat": 64000,
+    "deepseek-reasoner": 64000,
     ## OpenAI models: https://platform.openai.com/docs/models/overview
     # "o1-preview
     "chatgpt-4o-latest": 128000,

letta/functions/schema_generator.py

@@ -394,12 +394,12 @@ 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"]:
-        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.",
-        }
-        schema["parameters"]["required"].append("request_heartbeat")
+    # 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.",
+    }
+    schema["parameters"]["required"].append("request_heartbeat")
 
     return schema
 

letta/helpers/converters.py

@@ -0,0 +1,153 @@
+import base64
+from typing import Any, Dict, List, Optional, Union
+
+import numpy as np
+from openai.types.chat.chat_completion_message_tool_call import ChatCompletionMessageToolCall as OpenAIToolCall
+from openai.types.chat.chat_completion_message_tool_call import Function as OpenAIFunction
+from sqlalchemy import Dialect
+
+from letta.schemas.embedding_config import EmbeddingConfig
+from letta.schemas.enums import ToolRuleType
+from letta.schemas.llm_config import LLMConfig
+from letta.schemas.tool_rule import ChildToolRule, ConditionalToolRule, ContinueToolRule, InitToolRule, TerminalToolRule, ToolRule
+
+# --------------------------
+# LLMConfig Serialization
+# --------------------------
+
+
+def serialize_llm_config(config: Union[Optional[LLMConfig], Dict]) -> Optional[Dict]:
+    """Convert an LLMConfig object into a JSON-serializable dictionary."""
+    if config and isinstance(config, LLMConfig):
+        return config.model_dump()
+    return config
+
+
+def deserialize_llm_config(data: Optional[Dict]) -> Optional[LLMConfig]:
+    """Convert a dictionary back into an LLMConfig object."""
+    return LLMConfig(**data) if data else None
+
+
+# --------------------------
+# EmbeddingConfig Serialization
+# --------------------------
+
+
+def serialize_embedding_config(config: Union[Optional[EmbeddingConfig], Dict]) -> Optional[Dict]:
+    """Convert an EmbeddingConfig object into a JSON-serializable dictionary."""
+    if config and isinstance(config, EmbeddingConfig):
+        return config.model_dump()
+    return config
+
+
+def deserialize_embedding_config(data: Optional[Dict]) -> Optional[EmbeddingConfig]:
+    """Convert a dictionary back into an EmbeddingConfig object."""
+    return EmbeddingConfig(**data) if data else None
+
+
+# --------------------------
+# ToolRule Serialization
+# --------------------------
+
+
+def serialize_tool_rules(tool_rules: Optional[List[ToolRule]]) -> List[Dict[str, Any]]:
+    """Convert a list of ToolRules into a JSON-serializable format."""
+
+    if not tool_rules:
+        return []
+
+    data = [{**rule.model_dump(), "type": rule.type.value} for rule in tool_rules]  # Convert Enum to string for JSON compatibility
+
+    # Validate ToolRule structure
+    for rule_data in data:
+        if rule_data["type"] == ToolRuleType.constrain_child_tools.value and "children" not in rule_data:
+            raise ValueError(f"Invalid ToolRule serialization: 'children' field missing for rule {rule_data}")
+
+    return data
+
+
+def deserialize_tool_rules(data: Optional[List[Dict]]) -> List[Union[ChildToolRule, InitToolRule, TerminalToolRule, ConditionalToolRule]]:
+    """Convert a list of dictionaries back into ToolRule objects."""
+    if not data:
+        return []
+
+    return [deserialize_tool_rule(rule_data) for rule_data in data]
+
+
+def deserialize_tool_rule(data: Dict) -> Union[ChildToolRule, InitToolRule, TerminalToolRule, ConditionalToolRule, ContinueToolRule]:
+    """Deserialize a dictionary to the appropriate ToolRule subclass based on 'type'."""
+    rule_type = ToolRuleType(data.get("type"))
+
+    if rule_type == ToolRuleType.run_first:
+        return InitToolRule(**data)
+    elif rule_type == ToolRuleType.exit_loop:
+        return TerminalToolRule(**data)
+    elif rule_type == ToolRuleType.constrain_child_tools:
+        return ChildToolRule(**data)
+    elif rule_type == ToolRuleType.conditional:
+        return ConditionalToolRule(**data)
+    elif rule_type == ToolRuleType.continue_loop:
+        return ContinueToolRule(**data)
+    raise ValueError(f"Unknown ToolRule type: {rule_type}")
+
+
+# --------------------------
+# ToolCall Serialization
+# --------------------------
+
+
+def serialize_tool_calls(tool_calls: Optional[List[Union[OpenAIToolCall, dict]]]) -> List[Dict]:
+    """Convert a list of OpenAI ToolCall objects into JSON-serializable format."""
+    if not tool_calls:
+        return []
+
+    serialized_calls = []
+    for call in tool_calls:
+        if isinstance(call, OpenAIToolCall):
+            serialized_calls.append(call.model_dump())
+        elif isinstance(call, dict):
+            serialized_calls.append(call)  # Already a dictionary, leave it as-is
+        else:
+            raise TypeError(f"Unexpected tool call type: {type(call)}")
+
+    return serialized_calls
+
+
+def deserialize_tool_calls(data: Optional[List[Dict]]) -> List[OpenAIToolCall]:
+    """Convert a JSON list back into OpenAIToolCall objects."""
+    if not data:
+        return []
+
+    calls = []
+    for item in data:
+        func_data = item.pop("function", None)
+        tool_call_function = OpenAIFunction(**func_data) if func_data else None
+        calls.append(OpenAIToolCall(function=tool_call_function, **item))
+
+    return calls
+
+
+# --------------------------
+# Vector Serialization
+# --------------------------
+
+
+def serialize_vector(vector: Optional[Union[List[float], np.ndarray]]) -> Optional[bytes]:
+    """Convert a NumPy array or list into a base64-encoded byte string."""
+    if vector is None:
+        return None
+    if isinstance(vector, list):
+        vector = np.array(vector, dtype=np.float32)
+
+    return base64.b64encode(vector.tobytes())
+
+
+def deserialize_vector(data: Optional[bytes], dialect: Dialect) -> Optional[np.ndarray]:
+    """Convert a base64-encoded byte string back into a NumPy array."""
+    if not data:
+        return None
+
+    if dialect.name == "sqlite":
+        data = base64.b64decode(data)
+
+    return np.frombuffer(data, dtype=np.float32)

letta/helpers/tool_rule_solver.py

@@ -4,7 +4,7 @@ from typing import List, Optional, Union
 from pydantic import BaseModel, Field
 
 from letta.schemas.enums import ToolRuleType
-from letta.schemas.tool_rule import BaseToolRule, ChildToolRule, ConditionalToolRule, InitToolRule, TerminalToolRule
+from letta.schemas.tool_rule import BaseToolRule, ChildToolRule, ConditionalToolRule, ContinueToolRule, InitToolRule, TerminalToolRule
 
 
 class ToolRuleValidationError(Exception):
@@ -18,6 +18,9 @@ class ToolRulesSolver(BaseModel):
     init_tool_rules: List[InitToolRule] = Field(
         default_factory=list, description="Initial tool rules to be used at the start of tool execution."
     )
+    continue_tool_rules: List[ContinueToolRule] = Field(
+        default_factory=list, description="Continue tool rules to be used to continue tool execution."
+    )
     tool_rules: List[Union[ChildToolRule, ConditionalToolRule]] = Field(
         default_factory=list, description="Standard tool rules for controlling execution sequence and allowed transitions."
     )
@@ -43,6 +46,9 @@ class ToolRulesSolver(BaseModel):
             elif rule.type == ToolRuleType.exit_loop:
                 assert isinstance(rule, TerminalToolRule)
                 self.terminal_tool_rules.append(rule)
+            elif rule.type == ToolRuleType.continue_loop:
+                assert isinstance(rule, ContinueToolRule)
+                self.continue_tool_rules.append(rule)
 
     def update_tool_usage(self, tool_name: str):
         """Update the internal state to track the last tool called."""
@@ -80,6 +86,10 @@ class ToolRulesSolver(BaseModel):
         """Check if the tool has children tools"""
         return any(rule.tool_name == tool_name for rule in self.tool_rules)
 
+    def is_continue_tool(self, tool_name):
+        """Check if the tool is defined as a continue tool in the tool rules."""
+        return any(rule.tool_name == tool_name for rule in self.continue_tool_rules)
+
     def validate_conditional_tool(self, rule: ConditionalToolRule):
         """
         Validate a conditional tool rule

letta/llm_api/anthropic.py

@@ -519,6 +519,7 @@ def _prepare_anthropic_request(
     prefix_fill: bool = True,
     # if true, put COT inside the tool calls instead of inside the content
     put_inner_thoughts_in_kwargs: bool = False,
+    bedrock: bool = False,
 ) -> dict:
     """Prepare the request data for Anthropic API format."""
 
@@ -606,10 +607,11 @@ def _prepare_anthropic_request(
     # NOTE: cannot prefill with tools for opus:
     # Your API request included an `assistant` message in the final position, which would pre-fill the `assistant` response. When using tools with "claude-3-opus-20240229"
     if prefix_fill and not put_inner_thoughts_in_kwargs and "opus" not in data["model"]:
-        data["messages"].append(
-            # Start the thinking process for the assistant
-            {"role": "assistant", "content": f"<{inner_thoughts_xml_tag}>"},
-        )
+        if not bedrock:  # not support for bedrock
+            data["messages"].append(
+                # Start the thinking process for the assistant
+                {"role": "assistant", "content": f"<{inner_thoughts_xml_tag}>"},
+            )
 
     # Validate max_tokens
     assert "max_tokens" in data, data
@@ -651,13 +653,16 @@ def anthropic_bedrock_chat_completions_request(
     inner_thoughts_xml_tag: Optional[str] = "thinking",
 ) -> ChatCompletionResponse:
     """Make a chat completion request to Anthropic via AWS Bedrock."""
-    data = _prepare_anthropic_request(data, inner_thoughts_xml_tag)
+    data = _prepare_anthropic_request(data, inner_thoughts_xml_tag, bedrock=True)
 
     # Get the client
     client = get_bedrock_client()
 
     # Make the request
     try:
+        # bedrock does not support certain args
+        data["tool_choice"] = {"type": "any"}
+
         response = client.messages.create(**data)
         return convert_anthropic_response_to_chatcompletion(response=response, inner_thoughts_xml_tag=inner_thoughts_xml_tag)
     except PermissionDeniedError:

letta/llm_api/aws_bedrock.py

@@ -10,7 +10,7 @@ def has_valid_aws_credentials() -> bool:
     """
     Check if AWS credentials are properly configured.
     """
-    valid_aws_credentials = os.getenv("AWS_ACCESS_KEY_ID") and os.getenv("AWS_SECRET_ACCESS_KEY") and os.getenv("AWS_REGION")
+    valid_aws_credentials = os.getenv("AWS_ACCESS_KEY") and os.getenv("AWS_SECRET_ACCESS_KEY") and os.getenv("AWS_REGION")
     return valid_aws_credentials