edsl 0.1.37.dev3__py3-none-any.whl → 0.1.37.dev5__py3-none-any.whl

edsl/conversation/Conversation.py

@@ -1,238 +1,290 @@
-from collections import UserList
-import asyncio
-import inspect
-from typing import Optional, Callable
-from edsl import Agent, QuestionFreeText, Results, AgentList, ScenarioList, Scenario
-from edsl.questions import QuestionBase
-from edsl.results.Result import Result
-
-from edsl.data import Cache
-
-from edsl.conversation.next_speaker_utilities import (
-    default_turn_taking_generator,
-    speaker_closure,
-)
-
-
-class AgentStatement:
-    def __init__(self, statement: Result):
-        self.statement = statement
-
-    @property
-    def agent_name(self):
-        return self.statement["agent"]["name"]
-
-    def to_dict(self):
-        return self.statement.to_dict()
-
-    @classmethod
-    def from_dict(cls, data):
-        return cls(Result.from_dict(data))
-
-    @property
-    def text(self):
-        return self.statement["answer"]["dialogue"]
-
-
-class AgentStatements(UserList):
-    def __init__(self, data=None):
-        super().__init__(data)
-
-    @property
-    def transcript(self):
-        return [{s.agent_name: s.text} for s in self.data]
-
-    def to_dict(self):
-        return [d.to_dict() for d in self.data]
-
-    @classmethod
-    def from_dict(cls, data):
-        return cls([AgentStatement.from_dict(d) for d in data])
-
-
-class Conversation:
-    """A conversation between a list of agents. The first agent in the list is the first speaker.
-    After that, order is determined by the next_speaker function.
-    The question asked to each agent is determined by the next_statement_question.
-    """
-
-    def __init__(
-        self,
-        agent_list: AgentList,
-        max_turns: int = 20,
-        stopping_function: Optional[Callable] = None,
-        next_statement_question: Optional[QuestionBase] = None,
-        next_speaker_generator: Optional[Callable] = None,
-        verbose: bool = False,
-        conversation_index: Optional[int] = None,
-        cache=None,
-    ):
-        if cache is None:
-            self.cache = Cache()
-        else:
-            self.cache = cache
-
-        self.agent_list = agent_list
-        self.verbose = verbose
-        self.agent_statements = []
-        self._conversation_index = conversation_index
-
-        self.agent_statements = AgentStatements()
-
-        self.max_turns = max_turns
-
-        if next_statement_question is None:
-            self.next_statement_question = QuestionFreeText(
-                question_text="You are {{ agent_name }}. This is the converstaion so far: {{ conversation }}. What do you say next?",
-                question_name="dialogue",
-            )
-
-        # Determine how the next speaker is chosen
-        if next_speaker_generator is None:
-            func = default_turn_taking_generator
-        else:
-            func = next_speaker_generator
-
-        self.next_speaker = speaker_closure(
-            agent_list=self.agent_list, generator_function=func
-        )
-
-        # Determine when the conversation ends
-        if stopping_function is None:
-            self.stopping_function = lambda agent_statements: False
-        else:
-            self.stopping_function = stopping_function
-
-    async def continue_conversation(self, **kwargs) -> bool:
-        if len(self.agent_statements) >= self.max_turns:
-            return False
-
-        if inspect.iscoroutinefunction(self.stopping_function):
-            should_stop = await self.stopping_function(self.agent_statements, **kwargs)
-        else:
-            should_stop = self.stopping_function(self.agent_statements, **kwargs)
-
-        return not should_stop
-
-    def add_index(self, index) -> None:
-        self._conversation_index = index
-
-    @property
-    def conversation_index(self):
-        return self._conversation_index
-
-    def to_dict(self):
-        return {
-            "agent_list": self.agent_list.to_dict(),
-            "max_turns": self.max_turns,
-            "verbose": self.verbose,
-            "agent_statements": [d.to_dict() for d in self.agent_statements],
-            "conversation_index": self.conversation_index,
-        }
-
-    @classmethod
-    def from_dict(cls, data):
-        agent_list = AgentList.from_dict(data["agent_list"])
-        max_turns = data["max_turns"]
-        verbose = data["verbose"]
-        agent_statements = (AgentStatements.from_dict(data["agent_statements"]),)
-        conversation_index = data["conversation_index"]
-        return cls(
-            agent_list=agent_list,
-            max_turns=max_turns,
-            verbose=verbose,
-            results_data=agent_statements,
-            conversation_index=conversation_index,
-        )
-
-    def to_results(self):
-        return Results(data=[s.statement for s in self.agent_statements])
-
-    def summarize(self):
-        d = {
-            "num_agents": len(self.agent_list),
-            "max_turns": self.max_turns,
-            "conversation_index": self.conversation_index,
-            "transcript": self.to_results().select("agent_name", "dialogue").to_list(),
-            "number_of_agent_statements": len(self.agent_statements),
-        }
-        return Scenario(d)
-
-    async def get_next_statement(self, *, index, speaker, conversation):
-        q = self.next_statement_question
-        assert q.parameters == {"agent_name", "conversation"}, q.parameters
-        results = await q.run_async(
-            index=index,
-            conversation=conversation,
-            conversation_index=self.conversation_index,
-            agent_name=speaker.name,
-            agent=speaker,
-            just_answer=False,
-            cache=self.cache,
-            model=speaker.model,
-        )
-        return results[0]
-
-    def converse(self):
-        return asyncio.run(self._converse())
-
-    async def _converse(self):
-        i = 0
-        while await self.continue_conversation():
-            speaker = self.next_speaker()
-
-            next_statement = AgentStatement(
-                statement=await self.get_next_statement(
-                    index=i,
-                    speaker=speaker,
-                    conversation=self.agent_statements.transcript,
-                )
-            )
-            self.agent_statements.append(next_statement)
-            if self.verbose:
-                print(f"'{speaker.name}':{next_statement.text}")
-                print("\n")
-            i += 1
-
-
-class ConversationList:
-    """A collection of conversations to be run in parallel."""
-
-    def __init__(self, conversations: list[Conversation], cache=None):
-        self.conversations = conversations
-        for i, conversation in enumerate(self.conversations):
-            conversation.add_index(i)
-
-        if cache is None:
-            self.cache = Cache()
-        else:
-            self.cache = cache
-
-        for c in self.conversations:
-            c.cache = self.cache
-
-    async def run_conversations(self):
-        await asyncio.gather(*[c._converse() for c in self.conversations])
-
-    def run(self) -> None:
-        """Run all conversations in parallel"""
-        asyncio.run(self.run_conversations())
-
-    def to_dict(self) -> dict:
-        return {"conversations": c.to_dict() for c in self.conversations}
-
-    @classmethod
-    def from_dict(cls, data):
-        conversations = [Conversation.from_dict(d) for d in data["conversations"]]
-        return cls(conversations)
-
-    def to_results(self) -> Results:
-        """Return the results of all conversations as a single Results"""
-        first_convo = self.conversations[0]
-        results = first_convo.to_results()
-        for conv in self.conversations[1:]:
-            results += conv.to_results()
-        return results
-
-    def summarize(self) -> ScenarioList:
-        return ScenarioList([c.summarize() for c in self.conversations])
+from collections import UserList
+import asyncio
+import inspect
+from typing import Optional, Callable
+from edsl import Agent, QuestionFreeText, Results, AgentList, ScenarioList, Scenario
+from edsl.questions import QuestionBase
+from edsl.results.Result import Result
+from jinja2 import Template
+from edsl.data import Cache
+
+from edsl.conversation.next_speaker_utilities import (
+    default_turn_taking_generator,
+    speaker_closure,
+)
+
+
+class AgentStatement:
+    def __init__(self, statement: Result):
+        self.statement = statement
+
+    @property
+    def agent_name(self):
+        return self.statement["agent"]["name"]
+
+    def to_dict(self):
+        return self.statement.to_dict()
+
+    @classmethod
+    def from_dict(cls, data):
+        return cls(Result.from_dict(data))
+
+    @property
+    def text(self):
+        return self.statement["answer"]["dialogue"]
+
+
+class AgentStatements(UserList):
+    def __init__(self, data=None):
+        super().__init__(data)
+
+    @property
+    def transcript(self):
+        return [{s.agent_name: s.text} for s in self.data]
+
+    def to_dict(self):
+        return [d.to_dict() for d in self.data]
+
+    @classmethod
+    def from_dict(cls, data):
+        return cls([AgentStatement.from_dict(d) for d in data])
+
+
+class Conversation:
+    """A conversation between a list of agents. The first agent in the list is the first speaker.
+    After that, order is determined by the next_speaker function.
+    The question asked to each agent is determined by the next_statement_question.
+
+    If the user has passed in a "per_round_message_template", this will be displayed at the beginning of each round.
+    {{ round_message }} must be in the question_text.
+    """
+
+    def __init__(
+        self,
+        agent_list: AgentList,
+        max_turns: int = 20,
+        stopping_function: Optional[Callable] = None,
+        next_statement_question: Optional[QuestionBase] = None,
+        next_speaker_generator: Optional[Callable] = None,
+        verbose: bool = False,
+        per_round_message_template: Optional[str] = None,
+        conversation_index: Optional[int] = None,
+        cache=None,
+        disable_remote_inference=False,
+        default_model: Optional["LanguageModel"] = None,
+    ):
+        self.disable_remote_inference = disable_remote_inference
+        self.per_round_message_template = per_round_message_template
+
+        if cache is None:
+            self.cache = Cache()
+        else:
+            self.cache = cache
+
+        self.agent_list = agent_list
+
+        from edsl import Model
+
+        for agent in self.agent_list:
+            if not hasattr(agent, "model"):
+                if default_model is not None:
+                    agent.model = default_model
+                else:
+                    agent.model = Model()
+
+        self.verbose = verbose
+        self.agent_statements = []
+        self._conversation_index = conversation_index
+        self.agent_statements = AgentStatements()
+
+        self.max_turns = max_turns
+
+        if next_statement_question is None:
+            import textwrap
+
+            base_question = textwrap.dedent(
+                """\
+You are {{ agent_name }}. This is the conversation so far: {{ conversation }}
+{% if round_message is not none %}
+{{ round_message }}
+{% endif %}
+What do you say next?"""
+            )
+            self.next_statement_question = QuestionFreeText(
+                question_text=base_question,
+                question_name="dialogue",
+            )
+        else:
+            self.next_statement_question = next_statement_question
+            if (
+                per_round_message_template
+                and "{{ round_message }}" not in next_statement_question.question_text
+            ):
+                raise ValueError(
+                    "If you pass in a per_round_message_template, you must include {{ round_message }} in the question_text."
+                )
+
+        # Determine how the next speaker is chosen
+        if next_speaker_generator is None:
+            func = default_turn_taking_generator
+        else:
+            func = next_speaker_generator
+
+        # Choose the next speaker
+        self.next_speaker = speaker_closure(
+            agent_list=self.agent_list, generator_function=func
+        )
+
+        # Determine when the conversation ends
+        if stopping_function is None:
+            self.stopping_function = lambda agent_statements: False
+        else:
+            self.stopping_function = stopping_function
+
+    async def continue_conversation(self, **kwargs) -> bool:
+        if len(self.agent_statements) >= self.max_turns:
+            return False
+
+        if inspect.iscoroutinefunction(self.stopping_function):
+            should_stop = await self.stopping_function(self.agent_statements, **kwargs)
+        else:
+            should_stop = self.stopping_function(self.agent_statements, **kwargs)
+
+        return not should_stop
+
+    def add_index(self, index) -> None:
+        self._conversation_index = index
+
+    @property
+    def conversation_index(self):
+        return self._conversation_index
+
+    def to_dict(self):
+        return {
+            "agent_list": self.agent_list.to_dict(),
+            "max_turns": self.max_turns,
+            "verbose": self.verbose,
+            "agent_statements": [d.to_dict() for d in self.agent_statements],
+            "conversation_index": self.conversation_index,
+        }
+
+    @classmethod
+    def from_dict(cls, data):
+        agent_list = AgentList.from_dict(data["agent_list"])
+        max_turns = data["max_turns"]
+        verbose = data["verbose"]
+        agent_statements = (AgentStatements.from_dict(data["agent_statements"]),)
+        conversation_index = data["conversation_index"]
+        return cls(
+            agent_list=agent_list,
+            max_turns=max_turns,
+            verbose=verbose,
+            results_data=agent_statements,
+            conversation_index=conversation_index,
+        )
+
+    def to_results(self):
+        return Results(data=[s.statement for s in self.agent_statements])
+
+    def summarize(self):
+        d = {
+            "num_agents": len(self.agent_list),
+            "max_turns": self.max_turns,
+            "conversation_index": self.conversation_index,
+            "transcript": self.to_results().select("agent_name", "dialogue").to_list(),
+            "number_of_agent_statements": len(self.agent_statements),
+        }
+        return Scenario(d)
+
+    async def get_next_statement(self, *, index, speaker, conversation) -> "Result":
+        """Get the next statement from the speaker."""
+        q = self.next_statement_question
+        # assert q.parameters == {"agent_name", "conversation"}, q.parameters
+        from edsl import Scenario
+
+        if self.per_round_message_template is None:
+            round_message = None
+        else:
+            round_message = Template(self.per_round_message_template).render(
+                {"max_turns": self.max_turns, "current_turn": index}
+            )
+
+        s = Scenario(
+            {
+                "agent_name": speaker.name,
+                "conversation": conversation,
+                "conversation_index": self.conversation_index,
+                "index": index,
+                "round_message": round_message,
+            }
+        )
+        jobs = q.by(s).by(speaker).by(speaker.model)
+        jobs.show_prompts()
+        results = await jobs.run_async(
+            cache=self.cache, disable_remote_inference=self.disable_remote_inference
+        )
+        return results[0]
+
+    def converse(self):
+        return asyncio.run(self._converse())
+
+    async def _converse(self):
+        i = 0
+        while await self.continue_conversation():
+            speaker = self.next_speaker()
+
+            next_statement = AgentStatement(
+                statement=await self.get_next_statement(
+                    index=i,
+                    speaker=speaker,
+                    conversation=self.agent_statements.transcript,
+                )
+            )
+            self.agent_statements.append(next_statement)
+            if self.verbose:
+                print(f"'{speaker.name}':{next_statement.text}")
+                print("\n")
+            i += 1
+
+
+class ConversationList:
+    """A collection of conversations to be run in parallel."""
+
+    def __init__(self, conversations: list[Conversation], cache=None):
+        self.conversations = conversations
+        for i, conversation in enumerate(self.conversations):
+            conversation.add_index(i)
+
+        if cache is None:
+            self.cache = Cache()
+        else:
+            self.cache = cache
+
+        for c in self.conversations:
+            c.cache = self.cache
+
+    async def run_conversations(self):
+        await asyncio.gather(*[c._converse() for c in self.conversations])
+
+    def run(self) -> None:
+        """Run all conversations in parallel"""
+        asyncio.run(self.run_conversations())
+
+    def to_dict(self) -> dict:
+        return {"conversations": c.to_dict() for c in self.conversations}
+
+    @classmethod
+    def from_dict(cls, data):
+        conversations = [Conversation.from_dict(d) for d in data["conversations"]]
+        return cls(conversations)
+
+    def to_results(self) -> Results:
+        """Return the results of all conversations as a single Results"""
+        first_convo = self.conversations[0]
+        results = first_convo.to_results()
+        for conv in self.conversations[1:]:
+            results += conv.to_results()
+        return results
+
+    def summarize(self) -> ScenarioList:
+        return ScenarioList([c.summarize() for c in self.conversations])

edsl/conversation/car_buying.py

@@ -1,58 +1,58 @@
-from edsl import Agent, AgentList, QuestionFreeText
-from edsl import Cache
-from edsl.conversation.Conversation import Conversation, ConversationList
-
-a1 = Agent(
-    name="Alice",
-    traits={
-        "motivation": """
-    You are Alice. You want to buy a car. You are talking to Bob, a car salesman.
-    It is very important to you that the steering wheel not whiff out of the window while you are driving.
-    Also, the car can have no space for mother-in-law.
-    You brought your brother-in-law along, Paul, who you have an antagonistic relationship with.
-    """
-    },
-)
-a2 = Agent(
-    name="Bob",
-    traits={
-        "motivation": "You are Bob, a car salesman. You are talking to Alice, who wants to buy a car."
-    },
-)
-a3 = Agent(
-    name="Paul",
-    traits={
-        "motivation": "You are Paul, you are Alice's brother. You think her concerns are foolish and you are critical of her."
-    },
-)
-
-c1 = Conversation(agent_list=AgentList([a1, a3, a2]), max_turns=5, verbose=True)
-c2 = Conversation(agent_list=AgentList([a1, a2]), max_turns=5, verbose=True)
-
-c = Cache.load("car_talk.json.gz")
-# breakpoint()
-combo = ConversationList([c1, c2], cache=c)
-combo.run()
-results = combo.to_results()
-results.select("conversation_index", "index", "agent_name", "dialogue").print(
-    format="rich"
-)
-
-q = QuestionFreeText(
-    question_text="""This was a conversation about buying a car: {{ transcript }}. 
-                    Was a brand or style of car mentioned? If so, what was it?
-                    """,
-    question_name="car_brand",
-)
-
-from edsl import QuestionList
-
-q_actors = QuestionList(
-    question_text="""This was a conversation about buying a car: {{ transcript }}. 
-                    Who were the actors in the conversation?
-                    """,
-    question_name="actors",
-)
-
-transcript_analysis = q.add_question(q_actors).by(combo.summarize()).run()
-transcript_analysis.select("car_brand", "actors").print(format="rich")
+from edsl import Agent, AgentList, QuestionFreeText
+from edsl import Cache
+from edsl.conversation.Conversation import Conversation, ConversationList
+
+a1 = Agent(
+    name="Alice",
+    traits={
+        "motivation": """
+    You are Alice. You want to buy a car. You are talking to Bob, a car salesman.
+    It is very important to you that the steering wheel not whiff out of the window while you are driving.
+    Also, the car can have no space for mother-in-law.
+    You brought your brother-in-law along, Paul, who you have an antagonistic relationship with.
+    """
+    },
+)
+a2 = Agent(
+    name="Bob",
+    traits={
+        "motivation": "You are Bob, a car salesman. You are talking to Alice, who wants to buy a car."
+    },
+)
+a3 = Agent(
+    name="Paul",
+    traits={
+        "motivation": "You are Paul, you are Alice's brother. You think her concerns are foolish and you are critical of her."
+    },
+)
+
+c1 = Conversation(agent_list=AgentList([a1, a3, a2]), max_turns=5, verbose=True)
+c2 = Conversation(agent_list=AgentList([a1, a2]), max_turns=5, verbose=True)
+
+c = Cache.load("car_talk.json.gz")
+# breakpoint()
+combo = ConversationList([c1, c2], cache=c)
+combo.run()
+results = combo.to_results()
+results.select("conversation_index", "index", "agent_name", "dialogue").print(
+    format="rich"
+)
+
+q = QuestionFreeText(
+    question_text="""This was a conversation about buying a car: {{ transcript }}. 
+                    Was a brand or style of car mentioned? If so, what was it?
+                    """,
+    question_name="car_brand",
+)
+
+from edsl import QuestionList
+
+q_actors = QuestionList(
+    question_text="""This was a conversation about buying a car: {{ transcript }}. 
+                    Who were the actors in the conversation?
+                    """,
+    question_name="actors",
+)
+
+transcript_analysis = q.add_question(q_actors).by(combo.summarize()).run()
+transcript_analysis.select("car_brand", "actors").print(format="rich")