File size: 6,536 Bytes
318db6e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import concurrent.futures
import threading, math
import asyncio, json, os
from dotenv import load_dotenv
from llama_index.core import PromptTemplate
from llama_index.core.workflow import (
    Context,
    Workflow,
    StartEvent,
    StopEvent,
    step,
)
from workflow.events import (
    SafeStartEvent,
    RefuseEvent,
    TokenEvent,
    ControlEvent
)
from workflow.vllm_model import MyVllm
from workflow.modules import MySQLChatStore, ToyStatusStore, prGreen, prRed, prYellow
from prompts.default_prompts import(
    ALIGNMENT_PROMPT,
    REFUSE_PROMPT,
    FEMALE_ROLEPLAY_PROMPT,
    MALE_ROLEPLAY_PROMPT,
    TOY_CONTROL_PROMPT_TEST,
    TOY_CONTROL_PROMPT
)

REFUSE_INTENTS = [
    "medical advice", "Overdose medication", "child pornography", "self-harm", "political bias", "racial hate speech", "illegal drugs", "not harmful", "violent tendencies", "weaponry", "religious hate", "Theft", "Robbery", "Body Disposal", "Forgery", "Smuggling", "Money laundering", "Extortion", "Terrorism", "Explosion", "Cyberattack & Hacking", "illegal stalking", "Arms trafficking"
]
OPERATIONS = ["vibrate"]

class RolePlayWorkflow(Workflow):
    def __init__(
        self, 
        response_llm: MyVllm,
        chat_store: MySQLChatStore,
        toy_status_store: ToyStatusStore,
        sessionId: str,
        gender: str,
        toy_names: list[str] | None,
        *args,
        **kwargs
    ):
        super().__init__(*args, **kwargs)
        self.response_llm = response_llm
        self.chat_store = chat_store
        self.sessionId = sessionId
        self.chat_history = self.chat_store.get_chat_history(self.sessionId)
        self.gender = gender
        self.toy_names = toy_names
        self.toy_status_store = toy_status_store
        self.current_pattern = self.toy_status_store.get_latest(self.sessionId)["pattern"]
        self.retry_ct = 0

    @step
    async def censor(self, ctx: Context, ev: StartEvent) -> SafeStartEvent | RefuseEvent | StartEvent:
        # process llm output
        fmt_messages = ALIGNMENT_PROMPT.format_messages(
            user_input=ev.query,
            intent_labels=REFUSE_INTENTS
        )
        response = self.response_llm.chat(fmt_messages).message.content
        try:
            response = json.loads(response)
            intent = response["intent"]
            lang = response["language"]
            prYellow(f"language: {lang}")
            # 检测中文输入默认用英文回复
            if lang.lower() in ["zh", "chinese"]:
                lang = "english"
            await ctx.set("language", lang)
        except:
            if self.retry_ct < 3:
                self.retry_ct += 1
                return StartEvent(query=ev.query)
            return SafeStartEvent(query=ev.query)
        # judge
        if intent in ("not harmful", "BDSM content"):
            return SafeStartEvent(query=ev.query)
        prRed(f"refuse: {intent}")
        return RefuseEvent(lang=lang)

    @step
    async def refuse(self, ctx: Context, ev: RefuseEvent) -> StopEvent:
        response = self.response_llm.stream(REFUSE_PROMPT, language=ev.lang)
        response_str = ""
        for token in response:
            response_str += token
            content = json.dumps({"content": token})
            ctx.write_event_to_stream(TokenEvent(token=f"data:{content}\n\n"))
            await asyncio.sleep(0)
        ctx.write_event_to_stream(TokenEvent(token=f"data:[DONE]\n\n"))

        prRed(f"Response: {response_str}")
        return StopEvent(result="success")

    @step
    async def chat(self, ctx: Context, ev: SafeStartEvent) -> StopEvent:
        # preprocess chat history
        self.chat_store.add_message(self.sessionId, "user", ev.query)

        # generate response
        response_str = ""
        match self.gender:
            case "male":
                prompt = MALE_ROLEPLAY_PROMPT
            case "female":
                prompt = FEMALE_ROLEPLAY_PROMPT
        response = self.response_llm.stream(
            prompt,
            user_input=ev.query,
            chat_history=self.chat_history
        )
        for token in response:
            response_str += token
            content = json.dumps({"content": token})
            ctx.write_event_to_stream(TokenEvent(token=f"data:{content}\n\n"))
            await asyncio.sleep(0.005)
        ctx.write_event_to_stream(TokenEvent(token=f"data:[DONE]\n\n"))
        # update chat history
        t = threading.Thread(target=self.chat_store.add_message, args=(self.sessionId, "assistant", response_str))
        t.start()
        prGreen(f"Response:\n{response_str}")
        # control toy
        if self.toy_names:
            pattern = await self.control_toy(ev.query)
            ctx.write_event_to_stream(TokenEvent(token=f"data:{pattern}\n\n"))
        
        return StopEvent(result="success")

    async def control_toy(self, user_input:str):
        command = generate_command(
            self.response_llm,
            TOY_CONTROL_PROMPT,
            user_input=user_input,
            chat_history=self.chat_history,
            toy_status=self.current_pattern,
            available_operations=OPERATIONS
        )
        command_str = json.dumps(command)
        prGreen(command_str)
        # ctx.write_event_to_stream(TokenEvent(token=f"data:{command_str}\n\n"))
        # await asyncio.sleep(0)
        for toy_name in self.toy_names:
            t = threading.Thread(target=self.toy_status_store.update, args=(self.sessionId, command_str, toy_name))
            t.start()
        return command_str
    
def generate_command(llm: MyVllm, prompt: PromptTemplate, **kwagrs):
    """
    Format:
    {
        "operation":{
            "pattern":[
                {
                    "duration": 100,
                    "operation": int(level),
                },
                {
                    "duration": 100,
                    "operation": int(level),
                },
            ]
        }
    }
    """
    retry_ct = 0
    fmt_messages = prompt.format_messages(**kwagrs)
    while True:
        response = llm.chat(fmt_messages).message.content
        try:
            response = json.loads(response)
            break
        except:
            if retry_ct > 3:
                return "Failed to generate command"
            retry_ct += 1
            continue
    return response

def post_process_command(command: dict):
    total_time = sum([item["duration"] for item in command["pattern"]])
    mults = math.ceil(10000 / total_time)