聊天机器人 基于规则的到神经系统到法学执法代理

import re


class RulePattern:
    def __init__(self, pattern, response_template):
        self.regex = re.compile(pattern, re.IGNORECASE)
        self.template = response_template


PATTERNS = [
    RulePattern(r"my name is (\w+)", "Nice to meet you, {0}."),
    RulePattern(r"i (need|want) (.+)", "Why do you {0} {1}?"),
    RulePattern(r"i feel (.+)", "Why do you feel {0}?"),
    RulePattern(r"(.*)", "Tell me more about that."),
]


def rule_based_respond(user_input):
    for pattern in PATTERNS:
        m = pattern.regex.match(user_input.strip())
        if m:
            return pattern.template.format(*m.groups())
    return "I don't understand."

from sentence_transformers import SentenceTransformer
import numpy as np


FAQ = [
    ("how do i reset my password", "Go to Settings > Security > Reset Password."),
    ("how do i cancel my order", "Go to Orders, find the order, click Cancel."),
    ("what is your return policy", "30-day returns on unused items, original packaging."),
]


encoder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")
faq_questions = [q for q, _ in FAQ]
faq_embeddings = encoder.encode(faq_questions, normalize_embeddings=True)


def faq_respond(user_input, threshold=0.5):
    q_emb = encoder.encode([user_input], normalize_embeddings=True)[0]
    sims = faq_embeddings @ q_emb
    best = int(np.argmax(sims))
    if sims[best] < threshold:
        return None
    return FAQ[best][1]

from transformers import pipeline

chatbot = pipeline("text2text-generation", model="google/flan-t5-small")

response = chatbot("Respond politely to: Hi there!", max_new_tokens=40)
print(response[0]["generated_text"])

def agent_loop(user_message, tools, llm, max_steps=5):
    history = [{"role": "user", "content": user_message}]
    for _ in range(max_steps):
        response = llm(history, tools=tools)
        tool_call = response.get("tool_call")
        if tool_call:
            tool_name = tool_call.get("name")
            args = tool_call.get("arguments")
            if not isinstance(tool_name, str) or tool_name not in tools:
                history.append({"role": "assistant", "tool_call": tool_call})
                history.append({"role": "tool", "name": str(tool_name), "content": f"error: unknown tool {tool_name!r}"})
                continue
            if not isinstance(args, dict):
                history.append({"role": "assistant", "tool_call": tool_call})
                history.append({"role": "tool", "name": tool_name, "content": f"error: arguments must be a dict, got {type(args).__name__}"})
                continue
            fn = tools[tool_name]
            result = fn(**args)
            history.append({"role": "assistant", "tool_call": tool_call})
            history.append({"role": "tool", "name": tool_name, "content": result})
        else:
            return response["content"]
    return "I could not complete the task in the step budget."

def hybrid_chat(user_input):
    if is_destructive_action(user_input):
        return structured_flow(user_input)

    faq_answer = faq_respond(user_input, threshold=0.6)
    if faq_answer:
        return faq_answer

    return agent_loop(user_input, tools, llm)


def is_destructive_action(text):
    danger_words = ["delete", "cancel", "charge", "refund", "transfer"]
    return any(w in text.lower() for w in danger_words)

---
name: chatbot-architect
description: Design a chatbot stack for a given use case.
version: 1.0.0
phase: 5
lesson: 17
tags: [nlp, agents, chatbot]
---

Given a product context (user need, compliance constraints, available tools, data volume), output:

1. Architecture. Rule-based, retrieval, neural, LLM agent, or hybrid (specify which paths go where).
2. LLM choice if applicable. Name the model family (Claude, GPT-4, Llama-3.1, Mixtral). Match to tool-use quality and cost.
3. Grounding strategy. RAG sources, retrieval method (see lesson 14), tool contracts.
4. Evaluation plan. Task success rate, tool-call correctness, off-task rate, hallucination rate on held-out dialogs.

Refuse to recommend a pure-LLM agent for any destructive action (payments, account deletion, data modification) without a structured confirmation flow. Refuse to skip the prompt-injection audit if the agent has write access to anything.