591 lines
22 KiB
Python

from __future__ import annotations
from pydantic import BaseModel, Field # type: ignore
from typing import (
Literal,
get_args,
List,
AsyncGenerator,
TYPE_CHECKING,
Optional,
ClassVar,
Any,
TypeAlias,
Dict,
Tuple,
)
import json
import time
import inspect
from abc import ABC
import asyncio
from prometheus_client import Counter, Summary, CollectorRegistry # type: ignore
from ..setup_logging import setup_logging
logger = setup_logging()
# Only import Context for type checking
if TYPE_CHECKING:
from ..context import Context
from .types import agent_registry
from .. import defines
from ..message import Message, Tunables
from ..metrics import Metrics
from ..tools import TickerValue, WeatherForecast, AnalyzeSite, DateTime, llm_tools # type: ignore -- dynamically added to __all__
from ..conversation import Conversation
class LLMMessage(BaseModel):
role: str = Field(default="")
content: str = Field(default="")
tool_calls: Optional[List[Dict]] = Field(default={}, exclude=True)
class Agent(BaseModel, ABC):
"""
Base class for all agent types.
This class defines the common attributes and methods for all agent types.
"""
# Agent management with pydantic
agent_type: Literal["base"] = "base"
_agent_type: ClassVar[str] = agent_type # Add this for registration
# Tunables (sets default for new Messages attached to this agent)
tunables: Tunables = Field(default_factory=Tunables)
# Agent properties
system_prompt: str # Mandatory
conversation: Conversation = Conversation()
context_tokens: int = 0
context: Optional[Context] = Field(
default=None, exclude=True
) # Avoid circular reference, require as param, and prevent serialization
metrics: Metrics = Field(default_factory=Metrics, exclude=True)
# context_size is shared across all subclasses
_context_size: ClassVar[int] = int(defines.max_context * 0.5)
@property
def context_size(self) -> int:
return Agent._context_size
@context_size.setter
def context_size(self, value: int):
Agent._context_size = value
def set_optimal_context_size(
self, llm: Any, model: str, prompt: str, ctx_buffer=2048
) -> int:
# # Get more accurate token count estimate using tiktoken or similar
# response = llm.generate(
# model=model,
# prompt=prompt,
# options={
# "num_ctx": self.context_size,
# "num_predict": 0,
# } # Don't generate any tokens, just tokenize
# )
# # The prompt_eval_count gives you the token count of your input
# tokens = response.get("prompt_eval_count", 0)
# Most models average 1.3-1.5 tokens per word
word_count = len(prompt.split())
tokens = int(word_count * 1.4)
# Add buffer for safety
total_ctx = tokens + ctx_buffer
if total_ctx > self.context_size:
logger.info(
f"Increasing context size from {self.context_size} to {total_ctx}"
)
# Grow the context size if necessary
self.context_size = max(self.context_size, total_ctx)
# Use actual model maximum context size
return self.context_size
# Class and pydantic model management
def __init_subclass__(cls, **kwargs) -> None:
"""Auto-register subclasses"""
super().__init_subclass__(**kwargs)
# Register this class if it has an agent_type
if hasattr(cls, "agent_type") and cls.agent_type != Agent._agent_type:
agent_registry.register(cls.agent_type, cls)
# def __init__(self, *, context=context, **data):
# super().__init__(**data)
# self.set_context(context)
def model_dump(self, *args, **kwargs) -> Any:
# Ensure context is always excluded, even with exclude_unset=True
kwargs.setdefault("exclude", set())
if isinstance(kwargs["exclude"], set):
kwargs["exclude"].add("context")
elif isinstance(kwargs["exclude"], dict):
kwargs["exclude"]["context"] = True
return super().model_dump(*args, **kwargs)
@classmethod
def valid_agent_types(cls) -> set[str]:
"""Return the set of valid agent_type values."""
return set(get_args(cls.__annotations__["agent_type"]))
def set_context(self, context: Context):
object.__setattr__(self, "context", context)
# Agent methods
def get_agent_type(self):
return self._agent_type
async def prepare_message(self, message: Message) -> AsyncGenerator[Message, None]:
"""
Prepare message with context information in message.preamble
"""
logger.info(f"{self.agent_type} - {inspect.stack()[0].function}")
self.metrics.prepare_count.labels(agent=self.agent_type).inc()
with self.metrics.prepare_duration.labels(agent=self.agent_type).time():
if not self.context:
raise ValueError("Context is not set for this agent.")
# Generate RAG content if enabled, based on the content
rag_context = ""
if message.tunables.enable_rag and message.prompt:
# Gather RAG results, yielding each result
# as it becomes available
for message in self.context.generate_rag_results(message):
logger.info(f"RAG: {message.status} - {message.response}")
if message.status == "error":
yield message
return
if message.status != "done":
yield message
for rag in message.metadata.rag:
for doc in rag.documents:
rag_context += f"{doc}\n"
message.preamble = {}
if rag_context:
message.preamble["context"] = rag_context
if message.tunables.enable_context and self.context.user_resume:
message.preamble["resume"] = self.context.user_resume
message.system_prompt = self.system_prompt
message.status = "done"
yield message
return
async def process_tool_calls(
self,
llm: Any,
model: str,
message: Message,
tool_message: Any, # llama response message
messages: List[LLMMessage],
) -> AsyncGenerator[Message, None]:
logger.info(f"{self.agent_type} - {inspect.stack()[0].function}")
self.metrics.tool_count.labels(agent=self.agent_type).inc()
with self.metrics.tool_duration.labels(agent=self.agent_type).time():
if not self.context:
raise ValueError("Context is not set for this agent.")
if not message.metadata.tools:
raise ValueError("tools field not initialized")
tool_metadata = message.metadata.tools
tool_metadata["tool_calls"] = []
message.status = "tooling"
for i, tool_call in enumerate(tool_message.tool_calls):
arguments = tool_call.function.arguments
tool = tool_call.function.name
# Yield status update before processing each tool
message.response = (
f"Processing tool {i+1}/{len(tool_message.tool_calls)}: {tool}..."
)
yield message
logger.info(f"LLM - {message.response}")
# Process the tool based on its type
match tool:
case "TickerValue":
ticker = arguments.get("ticker")
if not ticker:
ret = None
else:
ret = TickerValue(ticker)
case "AnalyzeSite":
url = arguments.get("url")
question = arguments.get(
"question", "what is the summary of this content?"
)
# Additional status update for long-running operations
message.response = (
f"Retrieving and summarizing content from {url}..."
)
yield message
ret = await AnalyzeSite(
llm=llm, model=model, url=url, question=question
)
case "DateTime":
tz = arguments.get("timezone")
ret = DateTime(tz)
case "WeatherForecast":
city = arguments.get("city")
state = arguments.get("state")
message.response = (
f"Fetching weather data for {city}, {state}..."
)
yield message
ret = WeatherForecast(city, state)
case _:
ret = None
# Build response for this tool
tool_response = {
"role": "tool",
"content": json.dumps(ret),
"name": tool_call.function.name,
}
tool_metadata["tool_calls"].append(tool_response)
if len(tool_metadata["tool_calls"]) == 0:
message.status = "done"
yield message
return
message_dict = LLMMessage(
role=tool_message.get("role", "assistant"),
content=tool_message.get("content", ""),
tool_calls=[
{
"function": {
"name": tc["function"]["name"],
"arguments": tc["function"]["arguments"],
}
}
for tc in tool_message.tool_calls
],
)
messages.append(message_dict)
messages.extend(tool_metadata["tool_calls"])
message.status = "thinking"
message.response = "Incorporating tool results into response..."
yield message
# Decrease creativity when processing tool call requests
message.response = ""
start_time = time.perf_counter()
for response in llm.chat(
model=model,
messages=messages,
options={
**message.metadata.options,
},
stream=True,
):
# logger.info(f"LLM::Tools: {'done' if response.done else 'processing'} - {response.message}")
message.status = "streaming"
message.chunk = response.message.content
message.response += message.chunk
if not response.done:
yield message
if response.done:
self.collect_metrics(response)
message.metadata.eval_count += response.eval_count
message.metadata.eval_duration += response.eval_duration
message.metadata.prompt_eval_count += response.prompt_eval_count
message.metadata.prompt_eval_duration += response.prompt_eval_duration
self.context_tokens = (
response.prompt_eval_count + response.eval_count
)
message.status = "done"
yield message
end_time = time.perf_counter()
message.metadata.timers["llm_with_tools"] = end_time - start_time
return
def collect_metrics(self, response):
self.metrics.tokens_prompt.labels(agent=self.agent_type).inc(
response.prompt_eval_count
)
self.metrics.tokens_eval.labels(agent=self.agent_type).inc(response.eval_count)
async def generate_llm_response(
self, llm: Any, model: str, message: Message, temperature=0.7
) -> AsyncGenerator[Message, None]:
logger.info(f"{self.agent_type} - {inspect.stack()[0].function}")
self.metrics.generate_count.labels(agent=self.agent_type).inc()
with self.metrics.generate_duration.labels(agent=self.agent_type).time():
if not self.context:
raise ValueError("Context is not set for this agent.")
# Create a pruned down message list based purely on the prompt and responses,
# discarding the full preamble generated by prepare_message
messages: List[LLMMessage] = [
LLMMessage(role="system", content=message.system_prompt)
]
messages.extend(
[
item
for m in self.conversation
for item in [
LLMMessage(role="user", content=m.prompt.strip()),
LLMMessage(role="assistant", content=m.response.strip()),
]
]
)
# Only the actual user query is provided with the full context message
messages.append(
LLMMessage(role="user", content=message.context_prompt.strip())
)
# message.messages = messages
message.metadata.options = {
"seed": 8911,
"num_ctx": self.context_size,
"temperature": temperature, # Higher temperature to encourage tool usage
}
# Create a dict for storing various timing stats
message.metadata.timers = {}
use_tools = message.tunables.enable_tools and len(self.context.tools) > 0
message.metadata.tools = {
"available": llm_tools(self.context.tools),
"used": False,
}
tool_metadata = message.metadata.tools
if use_tools:
message.status = "thinking"
message.response = f"Performing tool analysis step 1/2..."
yield message
logger.info("Checking for LLM tool usage")
start_time = time.perf_counter()
# Tools are enabled and available, so query the LLM with a short context of messages
# in case the LLM did something like ask "Do you want me to run the tool?" and the
# user said "Yes" -- need to keep the context in the thread.
tool_metadata["messages"] = (
[{"role": "system", "content": self.system_prompt}] + messages[-6:]
if len(messages) >= 7
else messages
)
response = llm.chat(
model=model,
messages=tool_metadata["messages"],
tools=tool_metadata["available"],
options={
**message.metadata.options,
},
stream=False, # No need to stream the probe
)
self.collect_metrics(response)
end_time = time.perf_counter()
message.metadata.timers["tool_check"] = end_time - start_time
if not response.message.tool_calls:
logger.info("LLM indicates tools will not be used")
# The LLM will not use tools, so disable use_tools so we can stream the full response
use_tools = False
else:
tool_metadata["attempted"] = response.message.tool_calls
if use_tools:
logger.info("LLM indicates tools will be used")
# Tools are enabled and available and the LLM indicated it will use them
message.response = (
f"Performing tool analysis step 2/2 (tool use suspected)..."
)
yield message
logger.info(f"Performing LLM call with tools")
start_time = time.perf_counter()
response = llm.chat(
model=model,
messages=tool_metadata["messages"], # messages,
tools=tool_metadata["available"],
options={
**message.metadata.options,
},
stream=False,
)
self.collect_metrics(response)
end_time = time.perf_counter()
message.metadata.timers["non_streaming"] = end_time - start_time
if not response:
message.status = "error"
message.response = "No response from LLM."
yield message
return
if response.message.tool_calls:
tool_metadata["used"] = response.message.tool_calls
# Process all yielded items from the handler
start_time = time.perf_counter()
async for message in self.process_tool_calls(
llm=llm,
model=model,
message=message,
tool_message=response.message,
messages=messages,
):
if message.status == "error":
yield message
return
yield message
end_time = time.perf_counter()
message.metadata.timers["process_tool_calls"] = end_time - start_time
message.status = "done"
return
logger.info("LLM indicated tools will be used, and then they weren't")
message.response = response.message.content
message.status = "done"
yield message
return
# not use_tools
message.status = "thinking"
message.response = f"Generating response..."
yield message
# Reset the response for streaming
message.response = ""
start_time = time.perf_counter()
for response in llm.chat(
model=model,
messages=messages,
options={
**message.metadata.options,
},
stream=True,
):
if not response:
message.status = "error"
message.response = "No response from LLM."
yield message
return
message.status = "streaming"
message.chunk = response.message.content
message.response += message.chunk
if not response.done:
yield message
if response.done:
self.collect_metrics(response)
message.metadata.eval_count += response.eval_count
message.metadata.eval_duration += response.eval_duration
message.metadata.prompt_eval_count += response.prompt_eval_count
message.metadata.prompt_eval_duration += response.prompt_eval_duration
self.context_tokens = (
response.prompt_eval_count + response.eval_count
)
message.status = "done"
yield message
end_time = time.perf_counter()
message.metadata.timers["streamed"] = end_time - start_time
return
async def process_message(
self, llm: Any, model: str, message: Message
) -> AsyncGenerator[Message, None]:
logger.info(f"{self.agent_type} - {inspect.stack()[0].function}")
self.metrics.process_count.labels(agent=self.agent_type).inc()
with self.metrics.process_duration.labels(agent=self.agent_type).time():
if not self.context:
raise ValueError("Context is not set for this agent.")
logger.info(
"TODO: Implement delay queing; busy for same agent, otherwise return queue size and estimated wait time"
)
spinner: List[str] = ["\\", "|", "/", "-"]
tick: int = 0
while self.context.processing:
message.status = "waiting"
message.response = (
f"Busy processing another request. Please wait. {spinner[tick]}"
)
tick = (tick + 1) % len(spinner)
yield message
await asyncio.sleep(1) # Allow the event loop to process the write
self.context.processing = True
message.system_prompt = (
f"<|system|>\n{self.system_prompt.strip()}\n</|system|>"
)
message.context_prompt = ""
for p in message.preamble.keys():
message.context_prompt += (
f"\n<|{p}|>\n{message.preamble[p].strip()}\n</|{p}>\n\n"
)
message.context_prompt += f"{message.prompt}"
# Estimate token length of new messages
message.response = f"Optimizing context..."
message.status = "thinking"
yield message
message.context_size = self.set_optimal_context_size(
llm, model, prompt=message.context_prompt
)
message.response = f"Processing {'RAG augmented ' if message.metadata.rag else ''}query..."
message.status = "thinking"
yield message
async for message in self.generate_llm_response(
llm=llm, model=model, message=message
):
# logger.info(f"LLM: {message.status} - {f'...{message.response[-20:]}' if len(message.response) > 20 else message.response}")
if message.status == "error":
yield message
self.context.processing = False
return
yield message
# Done processing, add message to conversation
message.status = "done"
self.conversation.add(message)
self.context.processing = False
return
# Register the base agent
agent_registry.register(Agent._agent_type, Agent)