trustgraph/agent/agent-demo7
Cyber MacGeddon 4cc029f00c Agent infra
2024-11-09 23:08:43 +00:00

302 lines
7.5 KiB
Python
Executable file

#!/usr/bin/env python3
from trustgraph.clients.prompt_client import PromptClient
import json
import textwrap
import ibis
import textwrap
import time
import dataclasses
class Input:
state: dict
plan_state: str
plan: str
iteration: int
class Output:
state: dict
thought: str
next_state: str
plan: str
iteration: int
def output(text, prefix="> ", width=78):
out = textwrap.wrap(
text, initial_indent=prefix, subsequent_indent=prefix,
width=width
)
print("\n".join(out))
pulsar_host = "pulsar://localhost:6650"
prompt_client = PromptClient(pulsar_host=pulsar_host)
class State:
pass
class ActionWikipedia:
def __init__(self, prompt_client):
self.prompt_client = prompt_client
def act(self, state, topic):
pr = ibis.Template(
"Tell me what you think Wikipedia would say about {{topic}}"
)
input = pr.render({
"topic": topic,
})
resp = self.prompt_client.request(
"question",
{
"question": input
}
)
state.previous.append(
f"Question: {input.strip()}"
)
state.previous.append(
f"Answer: {resp.strip()}"
)
return state
class ActionCalculate:
def __init__(self, prompt_client):
self.prompt_client = prompt_client
def act(self, state, calc):
pr = ibis.Template(
"Compute this: {{computation}}. Just answer the question without explanation."
)
input = pr.render({
"computation": calc,
})
resp = self.prompt_client.request(
"question",
{
"question": input
}
)
print(resp)
state.previous.append(
"Calculate: " + calc.strip()
)
state.previous.append(
"Answer: " + resp.strip()
)
return state
class ActionAnswer:
def __init__(self, prompt_client):
self.prompt_client = prompt_client
def act(self, state, question):
pr = ibis.Template(
"Answer this: {{question}}"
)
input = pr.render({
"question": question,
})
resp = self.prompt_client.request(
"question",
{
"question": input
}
)
state.facts.append(
"Answer to " + input + ":\n" +
resp
)
return state
tools = {
"calculate": {
"description": "Takes a numeric computation and calculates the answer",
"implementation": ActionCalculate,
},
"cats-knowledge-store": {
"description": "Answer questions on Mark's cats using a knowledge store",
"implementation": ActionCalculate,
},
# "wikipedia": {
# "description": "Takes a query and looks it up on Wikipedia",
# "implementation": ActionWikipedia,
# },
# "answer": {
# "description": "Take a simple question and use the LLM to provide the answer",
# "implementation": ActionAnswer,
# },
# "space-shuttle-rag": {
# "description": "Take a question about space shuttles and answer it using the TrustGraph GraphRAG service",
# "implementation": ActionAnswer,
# }
}
class AgentManager:
determine_prompt = ibis.Template("""
You have access to the following tools:
{% for id, tool in tools.items() %}- tool-name: {{id}}
tool-description: {{tool.description}}
{% endfor %}
You operate in a loop. For each iteration of the loop, you take the
question and learnt knowledge. If the knowledge is enough to answer
the question, you will response with an answer. If the knowledge is not
enough, you will response with an action to be invoked to acquire more
knowledge.
Your output must ALWAYS be a well-formed JSON object. Output ONLY a JSON
object with no explanatory text or markup formatting. The JSON object
can have the following fields:
- thought: Mandatory. Your explanation of the current step, including the
goal and the reason for choosing this action. A string.
- action: An optional field, one of the tool-names above if you are using
an action.
- action-input: An optional field, input to the selected tool if you are using
an action.
- failure: a boolean. Set to true if you cannot proceed.
- final-answer: Your answer to the question as a string.
When you know the final answer to the original question, emit a 'final-answer'
JSON object with the following fields: thought and final-answer.
When you don't know the answer but know an action to move towards the
solution, emit an 'action' JSON object with the following fields:
thought, action, action-input
When you are stuck and aren't able to move torwards a solution, emit a
'failure' JSON object the following fields: thought, failure
{{previous}}""")
def __init__(self, tools, prompt_client):
self.tools = tools
self.prompt_client = prompt_client
def determine(self, state, thought=None):
input = __class__.determine_prompt.render({
"previous": "\n".join(state.previous),
"tools": self.tools,
})
print(input)
resp = self.prompt_client.request(
"question",
{
"question": input
}
)
resp = resp.replace("```json", "")
resp = resp.replace("```", "")
print(resp)
# print(resp)
resp = json.loads(resp)
# print(json.dumps(resp, indent=4))
return resp
def invoke(self, q, thought):
state = State()
state.question = q
state.previous = [
f"Question: {q}"
]
while True:
resp = self.determine(state)
print(resp)
if "thought" in resp:
if thought:
thought(resp["thought"])
state.previous.append(f"Thought: {resp['thought'].strip()}")
if "failure" in resp:
if resp["failure"]:
print("Failed")
return "Failed"
if "final-answer" in resp:
return resp["final-answer"]
if "action" not in resp:
raise RuntimeError("Didn't get final-answer/action response")
print(resp)
action = resp["action"]
print(action)
if action not in self.tools:
raise RuntimeError(f"Tool {action} not known")
if "implementation" not in self.tools[action]:
raise RuntimeError(f"Tool {action} not implemented")
if "action-input" in resp:
input = resp["action-input"]
else:
input = None
print("Action:", action)
print("Action input:", input)
print()
impl_class = self.tools[action]["implementation"]
impl = impl_class(self.prompt_client)
state = impl.act(state, input)
time.sleep(2)
def thought(t):
output("\U0001f914... " + t, prefix="? ")
print()
q = """Take the square root of 1600. That number is the number of a US
President. Who is that president and how do they relate to the space
shuttle programme?
"""
q = "How many eggs do I have if I have 3 eggs and Fred gives me some more eggs. The number of eggs he gives me is the square root of 64."
q = "Find out the behavioral nature of Mark's cats"
am = AgentManager(tools, prompt_client)
output("Q: " + q)
print()
resp = am.invoke(q, thought)
print(resp)