langroid 0.1.229__tar.gz → 0.1.231__tar.gz

{langroid-0.1.229 → langroid-0.1.231}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: langroid
-Version: 0.1.229
+Version: 0.1.231
 Summary: Harness LLMs with Multi-Agent Programming
 License: MIT
 Author: Prasad Chalasani

{langroid-0.1.229 → langroid-0.1.231}/langroid/agent/batch.py

@@ -31,6 +31,7 @@ def run_batch_task_gen(
     turns: int = -1,
     message: Optional[str] = None,
     handle_exceptions: bool = False,
+    max_cost: float = 0.0,
 ) -> list[U]:
     """
     Generate and run copies of a task async/concurrently one per item in `items` list.
@@ -50,6 +51,7 @@ def run_batch_task_gen(
         turns (int): number of turns to run, -1 for infinite
         message (Optional[str]): optionally overrides the console status messages
         handle_exceptions: bool: Whether to replace exceptions with outputs of None
+        max_cost: float: maximum cost to run the task (default 0.0 for unlimited)
 
     Returns:
         list[Any]: list of final results
@@ -62,7 +64,7 @@ def run_batch_task_gen(
             task_i.agent.llm.set_stream(False)
         task_i.agent.config.show_stats = False
 
-        result = await task_i.run_async(input, turns=turns)
+        result = await task_i.run_async(input, turns=turns, max_cost=max_cost)
         return result
 
     async def _do_all(
@@ -120,6 +122,7 @@ def run_batch_tasks(
     sequential: bool = True,
     batch_size: Optional[int] = None,
     turns: int = -1,
+    max_cost: float = 0.0,
 ) -> List[U]:
     """
     Run copies of `task` async/concurrently one per item in `items` list.
@@ -137,6 +140,7 @@ def run_batch_tasks(
         batch_size (Optional[int]): The number of tasks to run at a time,
             if None, unbatched
         turns (int): number of turns to run, -1 for infinite
+        max_cost: float: maximum cost to run the task (default 0.0 for unlimited)
 
     Returns:
         list[Any]: list of final results
@@ -151,6 +155,7 @@ def run_batch_tasks(
         batch_size,
         turns,
         message,
+        max_cost=max_cost,
     )
 
 

{langroid-0.1.229 → langroid-0.1.231}/langroid/agent/task.py

@@ -170,7 +170,7 @@ class Task:
                 agent.set_system_message(system_message)
             if user_message:
                 agent.set_user_message(user_message)
-
+        self.max_cost: float = 0
         self.logger: None | RichFileLogger = None
         self.tsv_logger: None | logging.Logger = None
         self.color_log: bool = False if settings.notebook else True
@@ -375,11 +375,13 @@ class Task:
         msg: Optional[str | ChatDocument] = None,
         turns: int = -1,
         caller: None | Task = None,
+        max_cost: float = 0,
     ) -> Optional[ChatDocument]:
         """Synchronous version of `run_async()`.
         See `run_async()` for details."""
         self.task_progress = False
         self.n_stalled_steps = 0
+        self.max_cost = max_cost
         assert (
             msg is None or isinstance(msg, str) or isinstance(msg, ChatDocument)
         ), f"msg arg in Task.run() must be None, str, or ChatDocument, not {type(msg)}"
@@ -418,6 +420,7 @@ class Task:
         msg: Optional[str | ChatDocument] = None,
         turns: int = -1,
         caller: None | Task = None,
+        max_cost: float = 0,
     ) -> Optional[ChatDocument]:
         """
         Loop over `step()` until task is considered done or `turns` is reached.
@@ -434,6 +437,7 @@ class Task:
             turns (int): number of turns to run the task for;
                 default is -1, which means run until task is done.
             caller (Task|None): the calling task, if any
+            max_cost (float): maximum cost allowed for the task (default 0 -> no limit)
 
         Returns:
             Optional[ChatDocument]: valid result of the task.
@@ -444,6 +448,8 @@ class Task:
         # message can be considered to be from the USER
         # (from the POV of this agent's LLM).
         self.task_progress = False
+        self.n_stalled_steps = 0
+        self.max_cost = max_cost
         if (
             isinstance(msg, ChatDocument)
             and msg.metadata.recipient != ""
@@ -819,6 +825,7 @@ class Task:
                 self.pending_message,
                 turns=actual_turns,
                 caller=self,
+                max_cost=self.max_cost,
             )
             result_str = str(ChatDocument.to_LLMMessage(result))
             maybe_tool = len(extract_top_level_json(result_str)) > 0
@@ -1051,6 +1058,16 @@ class Task:
             )
             return True
 
+        if self.max_cost > 0 and self.agent.llm is not None:
+            try:
+                if self.agent.llm.tot_tokens_cost()[1] > self.max_cost:
+                    logger.warning(
+                        f"Task {self.name} exceeded max cost {self.max_cost}; exiting."
+                    )
+                    return True
+            except Exception:
+                pass
+
         return (
             # no valid response from any entity/agent in current turn
             result is None

{langroid-0.1.229 → langroid-0.1.231}/langroid/language_models/base.py

@@ -449,6 +449,18 @@ class LanguageModel(ABC):
             s += f"{model}: {counter}\n"
         return s
 
+    @classmethod
+    def tot_tokens_cost(cls) -> Tuple[int, float]:
+        """
+        Return total tokens used and total cost across all models.
+        """
+        total_tokens = 0
+        total_cost = 0.0
+        for counter in cls.usage_cost_dict.values():
+            total_tokens += counter.total_tokens
+            total_cost += counter.cost
+        return total_tokens, total_cost
+
     def followup_to_standalone(
         self, chat_history: List[Tuple[str, str]], question: str
     ) -> str:

{langroid-0.1.229 → langroid-0.1.231}/langroid/language_models/openai_gpt.py

@@ -49,7 +49,7 @@ from langroid.language_models.utils import (
     retry_with_exponential_backoff,
 )
 from langroid.utils.configuration import settings
-from langroid.utils.constants import NO_ANSWER, Colors
+from langroid.utils.constants import Colors
 from langroid.utils.system import friendly_error
 
 logging.getLogger("openai").setLevel(logging.ERROR)
@@ -847,9 +847,9 @@ class OpenAIGPT(LanguageModel):
         try:
             return self._generate(prompt, max_tokens)
         except Exception as e:
-            # capture exceptions not handled by retry, so we don't crash
+            # log and re-raise exception
             logging.error(friendly_error(e, "Error in OpenAIGPT.generate: "))
-            return LLMResponse(message=NO_ANSWER, cached=False)
+            raise e
 
     def _generate(self, prompt: str, max_tokens: int) -> LLMResponse:
         if self.config.use_chat_for_completion:
@@ -917,9 +917,9 @@ class OpenAIGPT(LanguageModel):
         try:
             return await self._agenerate(prompt, max_tokens)
         except Exception as e:
-            # capture exceptions not handled by retry, so we don't crash
+            # log and re-raise exception
             logging.error(friendly_error(e, "Error in OpenAIGPT.agenerate: "))
-            return LLMResponse(message=NO_ANSWER, cached=False)
+            raise e
 
     async def _agenerate(self, prompt: str, max_tokens: int) -> LLMResponse:
         # note we typically will not have self.config.stream = True
@@ -1019,9 +1019,9 @@ class OpenAIGPT(LanguageModel):
         try:
             return self._chat(messages, max_tokens, functions, function_call)
         except Exception as e:
-            # capture exceptions not handled by retry, so we don't crash
+            # log and re-raise exception
             logging.error(friendly_error(e, "Error in OpenAIGPT.chat: "))
-            return LLMResponse(message=NO_ANSWER, cached=False)
+            raise e
 
     async def achat(
         self,
@@ -1072,9 +1072,9 @@ class OpenAIGPT(LanguageModel):
             result = await self._achat(messages, max_tokens, functions, function_call)
             return result
         except Exception as e:
-            # capture exceptions not handled by retry, so we don't crash
+            # log and re-raise exception
             logging.error(friendly_error(e, "Error in OpenAIGPT.achat: "))
-            return LLMResponse(message=NO_ANSWER, cached=False)
+            raise e
 
     @retry_with_exponential_backoff
     def _chat_completions_with_backoff(self, **kwargs):  # type: ignore

{langroid-0.1.229 → langroid-0.1.231}/langroid/language_models/utils.py

@@ -62,6 +62,7 @@ def retry_with_exponential_backoff(
                 if num_retries > max_retries:
                     raise Exception(
                         f"Maximum number of retries ({max_retries}) exceeded."
+                        f" Last error: {e}."
                     )
 
                 # Increment the delay

{langroid-0.1.229 → langroid-0.1.231}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "langroid"
-version = "0.1.229"
+version = "0.1.231"
 description = "Harness LLMs with Multi-Agent Programming"
 authors = ["Prasad Chalasani <pchalasani@gmail.com>"]
 readme = "README.md"