trustgraph/trustgraph-flow/trustgraph/agent/react/agent_manager.py
cybermaggedon b1cc724f7d
Streaming LLM part 2 (#567)
* Updates for agent API with streaming support

* Added tg-dump-queues tool to dump Pulsar queues to a log

* Updated tg-invoke-agent, incremental output

* Queue dumper CLI - might be useful for debug

* Updating for tests
2025-11-26 15:16:17 +00:00

386 lines
14 KiB
Python

import logging
import json
import re
import asyncio
from . types import Action, Final
logger = logging.getLogger(__name__)
class AgentManager:
def __init__(self, tools, additional_context=None):
self.tools = tools
self.additional_context = additional_context
def parse_react_response(self, text):
"""Parse text-based ReAct response format.
Expected format:
Thought: [reasoning about what to do next]
Action: [tool_name]
Args: {
"param": "value"
}
OR
Thought: [reasoning about the final answer]
Final Answer: [the answer]
"""
if not isinstance(text, str):
raise ValueError(f"Expected string response, got {type(text)}")
# Remove any markdown code blocks that might wrap the response
text = re.sub(r'^```[^\n]*\n', '', text.strip())
text = re.sub(r'\n```$', '', text.strip())
lines = text.strip().split('\n')
thought = None
action = None
args = None
final_answer = None
i = 0
while i < len(lines):
line = lines[i].strip()
# Parse Thought
if line.startswith("Thought:"):
thought = line[8:].strip()
# Handle multi-line thoughts
i += 1
while i < len(lines):
next_line = lines[i].strip()
if next_line.startswith(("Action:", "Final Answer:", "Args:")):
break
thought += " " + next_line
i += 1
continue
# Parse Final Answer
if line.startswith("Final Answer:"):
final_answer = line[13:].strip()
# Handle multi-line final answers (including JSON)
i += 1
# Check if the answer might be JSON
if final_answer.startswith('{') or (i < len(lines) and lines[i].strip().startswith('{')):
# Collect potential JSON answer
json_text = final_answer if final_answer.startswith('{') else ""
brace_count = json_text.count('{') - json_text.count('}')
while i < len(lines) and (brace_count > 0 or not json_text):
current_line = lines[i].strip()
if current_line.startswith(("Thought:", "Action:")) and brace_count == 0:
break
json_text += ("\n" if json_text else "") + current_line
brace_count += current_line.count('{') - current_line.count('}')
i += 1
# Try to parse as JSON
# try:
# final_answer = json.loads(json_text)
# except json.JSONDecodeError:
# # Not valid JSON, treat as regular text
# final_answer = json_text
final_answer = json_text
else:
# Regular text answer
while i < len(lines):
next_line = lines[i].strip()
if next_line.startswith(("Thought:", "Action:")):
break
final_answer += " " + next_line
i += 1
# If we have a final answer, return Final object
return Final(
thought=thought or "",
final=final_answer
)
# Parse Action
if line.startswith("Action:"):
action = line[7:].strip()
# Get rid of quotation prefix/suffix if present
while action and action[0] == '"':
action = action[1:]
while action and action[-1] == '"':
action = action[:-1]
# Parse Args
if line.startswith("Args:"):
# Check if JSON starts on the same line
args_on_same_line = line[5:].strip()
if args_on_same_line:
args_text = args_on_same_line
brace_count = args_on_same_line.count('{') - args_on_same_line.count('}')
else:
args_text = ""
brace_count = 0
# Collect all lines that form the JSON arguments
i += 1
started = bool(args_on_same_line and '{' in args_on_same_line)
while i < len(lines) and (not started or brace_count > 0):
current_line = lines[i]
args_text += ("\n" if args_text else "") + current_line
# Count braces to determine when JSON is complete
for char in current_line:
if char == '{':
brace_count += 1
started = True
elif char == '}':
brace_count -= 1
# If we've started and braces are balanced, we're done
if started and brace_count == 0:
break
i += 1
# Parse the JSON arguments
try:
args = json.loads(args_text.strip())
except json.JSONDecodeError as e:
logger.error(f"Failed to parse JSON arguments: {args_text}")
raise ValueError(f"Invalid JSON in Args: {e}")
i += 1
# If we have an action, return Action object
if action:
return Action(
thought=thought or "",
name=action,
arguments=args or {},
observation=""
)
# If we only have a thought but no action or final answer
if thought and not action and not final_answer:
raise ValueError(f"Response has thought but no action or final answer: {text}")
raise ValueError(f"Could not parse response: {text}")
async def reason(self, question, history, context, streaming=False, think=None, observe=None, answer=None):
logger.debug(f"calling reason: {question}")
tools = self.tools
logger.debug("in reason")
logger.debug(f"tools: {tools}")
tool_names = ",".join([
t for t in self.tools.keys()
])
logger.debug(f"Tool names: {tool_names}")
variables = {
"question": question,
"tools": [
{
"name": tool.name,
"description": tool.description,
"arguments": [
{
"name": arg.name,
"type": arg.type,
"description": arg.description
}
for arg in tool.arguments
]
}
for tool in self.tools.values()
],
"context": self.additional_context,
"question": question,
"tool_names": tool_names,
"history": [
{
"thought": h.thought,
"action": h.name,
"arguments": h.arguments,
"observation": h.observation,
}
for h in history
]
}
logger.debug(f"Variables: {json.dumps(variables, indent=4)}")
logger.info(f"prompt: {variables}")
logger.info(f"DEBUG: streaming={streaming}, think={think is not None}")
# Streaming path - use StreamingReActParser
if streaming and think:
logger.info("DEBUG: Entering streaming path")
from .streaming_parser import StreamingReActParser
logger.info("DEBUG: Creating StreamingReActParser")
# Collect chunks to send via async callbacks
thought_chunks = []
answer_chunks = []
# Create parser with synchronous callbacks that just collect chunks
parser = StreamingReActParser(
on_thought_chunk=lambda chunk: thought_chunks.append(chunk),
on_answer_chunk=lambda chunk: answer_chunks.append(chunk),
)
logger.info("DEBUG: StreamingReActParser created")
# Create async chunk callback that feeds parser and sends collected chunks
async def on_chunk(text):
logger.info(f"DEBUG: on_chunk called with {len(text)} chars")
# Track what we had before
prev_thought_count = len(thought_chunks)
prev_answer_count = len(answer_chunks)
# Feed the parser (synchronous)
logger.info(f"DEBUG: About to call parser.feed")
parser.feed(text)
logger.info(f"DEBUG: parser.feed returned")
# Send any new thought chunks
for i in range(prev_thought_count, len(thought_chunks)):
logger.info(f"DEBUG: Sending thought chunk {i}")
await think(thought_chunks[i])
# Send any new answer chunks
for i in range(prev_answer_count, len(answer_chunks)):
logger.info(f"DEBUG: Sending answer chunk {i}")
if answer:
await answer(answer_chunks[i])
else:
await think(answer_chunks[i])
logger.info("DEBUG: Getting prompt-request client from context")
client = context("prompt-request")
logger.info(f"DEBUG: Got client: {client}")
logger.info("DEBUG: About to call agent_react with streaming=True")
# Get streaming response
response_text = await client.agent_react(
variables=variables,
streaming=True,
chunk_callback=on_chunk
)
logger.info(f"DEBUG: agent_react returned, got {len(response_text) if response_text else 0} chars")
# Finalize parser
logger.info("DEBUG: Finalizing parser")
parser.finalize()
logger.info("DEBUG: Parser finalized")
# Get result
logger.info("DEBUG: Getting result from parser")
result = parser.get_result()
if result is None:
raise RuntimeError("Parser failed to produce a result")
logger.info(f"Parsed result: {result}")
return result
else:
logger.info("DEBUG: Entering NON-streaming path")
# Non-streaming path - get complete text and parse
logger.info("DEBUG: Getting prompt-request client from context")
client = context("prompt-request")
logger.info(f"DEBUG: Got client: {client}")
logger.info("DEBUG: About to call agent_react with streaming=False")
response_text = await client.agent_react(
variables=variables,
streaming=False
)
logger.info(f"DEBUG: agent_react returned, got response")
logger.debug(f"Response text:\n{response_text}")
logger.info(f"response: {response_text}")
# Parse the text response
try:
result = self.parse_react_response(response_text)
logger.info(f"Parsed result: {result}")
return result
except ValueError as e:
logger.error(f"Failed to parse response: {e}")
# Try to provide a helpful error message
logger.error(f"Response was: {response_text}")
raise RuntimeError(f"Failed to parse agent response: {e}")
async def react(self, question, history, think, observe, context, streaming=False, answer=None):
logger.info(f"question: {question}")
act = await self.reason(
question = question,
history = history,
context = context,
streaming = streaming,
think = think,
observe = observe,
answer = answer,
)
logger.info(f"act: {act}")
if isinstance(act, Final):
# In non-streaming mode, send complete thought
# In streaming mode, thoughts were already sent as chunks
if not streaming:
await think(act.thought)
return act
else:
# In non-streaming mode, send complete thought
# In streaming mode, thoughts were already sent as chunks
if not streaming:
await think(act.thought)
logger.debug(f"ACTION: {act.name}")
logger.debug(f"Tools: {self.tools.keys()}")
if act.name in self.tools:
action = self.tools[act.name]
else:
logger.debug(f"Tools: {self.tools}")
raise RuntimeError(f"No action for {act.name}!")
logger.debug(f"TOOL>>> {act}")
resp = await action.implementation(context).invoke(
**act.arguments
)
if isinstance(resp, str):
resp = resp.strip()
else:
resp = str(resp)
resp = resp.strip()
logger.info(f"resp: {resp}")
await observe(resp)
act.observation = resp
logger.info(f"iter: {act}")
return act