add and apply llm intent classifier

Dockerfile

@@ -19,6 +19,9 @@ RUN rasa train
 COPY requirements.txt /app/requirements.txt
 RUN pip install --no-cache-dir -r requirements.txt
 
+RUN pip install --no-cache-dir numpy==1.19.5
+RUN pip show numpy
+
 # 11) Copy your Gradio wrapper
 COPY app.py /app/app.py
 

classifier/classifier.py

@@ -0,0 +1,69 @@
+import logging
+from fastapi import FastAPI
+from pydantic import BaseModel
+from typing import List
+import os
+from string import Formatter
+
+import os
+
+import outlines
+from outlines.models import openai
+from outlines.generate import choice
+
+# Configure logger
+tools = logging.getLogger("classifier")
+tools.setLevel(logging.DEBUG)
+ch = logging.StreamHandler()
+ch.setLevel(logging.DEBUG)
+formatter = logging.Formatter("%(levelname)s:%(name)s:%(message)s")
+ch.setFormatter(formatter)
+tools.addHandler(ch)
+
+# Configure logger
+logging.basicConfig(
+    format="%(asctime)s %(levelname)s:%(name)s: %(message)s",
+    level=logging.DEBUG,
+)
+logger = logging.getLogger("classifier")
+
+app = FastAPI()
+
+# Pydantic model for incoming requests; prompt_template added
+class Req(BaseModel):
+    message: str
+    model_name: str
+    base_url: str
+    class_set: List[str]
+    prompt_template: str  # template with {message} placeholder
+
+class Resp(BaseModel):
+    result: str
+
+# Helper for safe formatting of {message} only
+class SafeFormatDict(dict):
+    def __missing__(self, key):
+        return '{' + key + '}'
+
+@app.post("/classify", response_model=Resp)
+def classify(req: Req):
+    logger.debug(f"Received request args: {req.dict()}")
+
+    prompt = req.prompt_template.replace("{message}", req.message)
+    logger.debug(f"Rendered prompt: {prompt!r}")
+
+    api_key = os.getenv("TOGETHERAI_API_KEY")
+    logger.debug(f"Using API_KEY: {'set' if api_key else 'missing'}")
+    llm = openai(req.model_name, api_key=api_key, base_url=req.base_url)
+    clf = choice(llm, req.class_set)
+    logger.debug(f"Choice classifier created with labels: {req.class_set}")
+
+    try:
+        result = clf(prompt)
+        # If it's a coroutine, run it; otherwise use result
+        logger.debug(f"Classifier returned: {result}")
+    except Exception as e:
+        result = req.class_set[-1]
+        logger.error(f"Classification error: {e}. Falling back to: {result}")
+
+    return Resp(result=result)

classifier/requirements.txt

@@ -0,0 +1,10 @@
+fastapi
+uvicorn
+outlines==0.2.1
+
+# Force CPU-only torch
+torch==2.0.1+cpu
+# (and any other torch-based libs you need, also +cpu)
+openai
+# Tell pip where to find the +cpu variants
+--extra-index-url https://download.pytorch.org/whl/cpu

config.yml

@@ -6,7 +6,66 @@ pipeline:
 # # No configuration for the NLU pipeline was provided. The following default pipeline was used to train your model.
 # # If you'd like to customize it, uncomment and adjust the pipeline.
 # # See https://rasa.com/docs/rasa/tuning-your-model for more information.
-- name: "KeywordIntentClassifier"
+#- name: "KeywordIntentClassifier"
+- name: custom_components.llm_intent_classifier_client.LlmIntentClassifier
+  # these two match the `defaults` keys in your component
+  model_name: "Qwen/Qwen2.5-7B-Instruct-Turbo"
+  base_url: "https://api.together.xyz/v1"
+  class_set:
+    - ReversePhoneLookup
+    - CallReservationHotline
+    - out_of_scope
+  # Override the built-in prompt template:
+  prompt_template: |
+    你是一個專門分類醫院電話總機使用者訊息的代理，需將訊息分為「ReversePhoneLookup」、「CallReservationHotline」或「out_of_scope」三類。
+
+    **定義**：
+    - 「ReversePhoneLookup」：病人詢問未接來電相關的訊息，通常是醫院曾撥打電話給病人但病人未接到，病人回電詢問來電者身份或來電目的。包含「未接」「錯過」「漏接」「誰打來」「什麼事」等關鍵詞。
+
+    - 「CallReservationHotline」：病人需要進行掛號預約或取消看診相關的訊息，包含預約門診、取消預約、更改看診時間、掛號等醫療預約相關事宜。
+
+    - 「out_of_scope」：不屬於上述兩類的訊息，包括：
+      - 詢問醫院一般資訊（營業時間、地址、科別等）
+      - 查詢檢查報告
+      - 轉接特定科別或醫師
+      - 其他非預約且非未接來電查詢的事宜
+
+    **示例**：
+    - ReversePhoneLookup：
+      - 「我剛才有未接來電，想知道是誰打來的？」
+      - 「請問剛才是醫院打電話給我嗎？有什麼事嗎？」
+      - 「我錯過了一通電話，可以告訴我是什麼事情嗎？」
+      - 「有人打電話給我但我沒接到，請問是關於什麼的？」
+      - 「我看到有未接來電，請問找我有什麼事？」
+
+    - CallReservationHotline：
+      - 「我想要預約心臟科的門診。」
+      - 「請幫我掛號看眼科。」
+      - 「我要取消明天下午的預約。」
+      - 「可以幫我更改看診時間嗎？」
+      - 「我想預約健康檢查。」
+      - 「需要重新安排我的復診時間。」
+
+    - out_of_scope：
+      - 「請問醫院的營業時間是什麼時候？」
+      - 「我想查詢我的檢查報告結果。」
+      - 「請幫我轉接腸胃科。」
+      - 「請問醫院地址在哪裡？」
+      - 「我想詢問住院相關事宜。」
+      - 「請問有哪些科別？」
+
+    **指令**：
+    - 請專注於訊息的主要意圖
+    - 若訊息涉及未接來電、錯過來電或詢問來電者身份，歸類為「ReversePhoneLookup」
+    - 若訊息涉及預約、掛號、取消或更改看診時間，歸類為「CallReservationHotline」
+    - 其他查詢、轉接或一般資訊需求歸類為「out_of_scope」
+
+    你必須僅回傳以下三種 JSON 物件之一：
+    {{"result": "ReversePhoneLookup"}} 或 {{"result": "CallReservationHotline"}} 或 {{"result": "out_of_scope"}}
+
+    請勿添加任何解釋或其他內容。
+
+    訊息：{message}
 - name: FallbackClassifier
   threshold: 0.7
 

custom_components/llm_intent_classifier_client.py

@@ -0,0 +1,114 @@
+import logging
+import requests
+from typing import Any, Dict, List, Optional, Text
+
+from rasa.nlu.classifiers.classifier import IntentClassifier
+from rasa.shared.nlu.constants import TEXT, INTENT
+from rasa.nlu.config import RasaNLUModelConfig
+from rasa.shared.nlu.training_data.training_data import TrainingData
+from rasa.shared.nlu.training_data.message import Message
+from rasa.nlu.model import Metadata
+
+logger = logging.getLogger(__name__)
+
+
+class LlmIntentClassifier(IntentClassifier):
+    """Delegates intent classification to an external HTTP micro-service."""
+
+    name = "LlmIntentClassifier"
+    defaults = {
+        "classifier_url": "http://classifier:8000/classify",
+        "timeout": 5.0,
+        "model_name": None,
+        "base_url": None,
+        "class_set": [],
+        "prompt_template": None,
+    }
+
+    def __init__(
+        self,
+        component_config: Optional[Dict[Text, Any]] = None,
+    ) -> None:
+        super().__init__(component_config or {})
+        self.url: str = self.component_config.get("classifier_url")
+        self.timeout: float = float(self.component_config.get("timeout"))
+        self.model_name: Optional[Text] = self.component_config.get("model_name")
+        self.base_url: Optional[Text] = self.component_config.get("base_url")
+        self.class_set: List[Text] = self.component_config.get("class_set", [])
+        self.prompt_template: Optional[Text] = self.component_config.get("prompt_template")
+
+        # Validate required configuration
+        missing: List[str] = []
+        if not self.model_name:
+            missing.append("model_name")
+        if not self.base_url:
+            missing.append("base_url")
+        if not self.class_set:
+            missing.append("class_set")
+        if not self.prompt_template:
+            missing.append("prompt_template")
+        if missing:
+            raise ValueError(
+                f"Missing configuration for {', '.join(missing)} in LlmIntentClassifier"
+            )
+
+    def train(
+        self,
+        training_data: TrainingData,
+        config: Optional[RasaNLUModelConfig] = None,
+        **kwargs: Any,
+    ) -> None:
+        # No local training; this uses a remote service
+        pass
+
+    def process(self, message: Message, **kwargs: Any) -> None:
+        text: Optional[Text] = message.get(TEXT)
+        intent_name: Optional[Text] = None
+        confidence: float = 0.0
+
+        if text:
+            payload: Dict[str, Any] = {
+                "message": text,
+                "model_name": self.model_name,
+                "base_url": self.base_url,
+                "class_set": self.class_set,
+                "prompt_template": self.prompt_template,
+            }
+            try:
+                resp = requests.post(self.url, json=payload, timeout=self.timeout)
+                resp.raise_for_status()
+                result = resp.json().get("result")
+                if isinstance(result, str):
+                    intent_name = result
+                    confidence = 1.0
+            except Exception as e:
+                logger.warning(f"LlmIntentClassifier HTTP error: {e}")
+
+        message.set(INTENT, {"name": intent_name, "confidence": confidence}, add_to_output=True)
+
+    def persist(
+        self,
+        file_name: Text,
+        model_dir: Text,
+    ) -> Optional[Dict[Text, Any]]:
+        # Save configuration so it can be reloaded
+        return {
+            "classifier_url": self.url,
+            "timeout": self.timeout,
+            "model_name": self.model_name,
+            "base_url": self.base_url,
+            "class_set": self.class_set,
+            "prompt_template": self.prompt_template,
+        }
+
+    @classmethod
+    def load(
+        cls,
+        meta: Dict[Text, Any],
+        model_dir: Text,
+        model_metadata: Metadata = None,
+        cached_component: Optional["LlmIntentClassifier"] = None,
+        **kwargs: Any,
+    ) -> "LlmIntentClassifier":
+        # meta contains saved configuration
+        return cls(meta)

requirements.txt

@@ -1,7 +1,5 @@
 # Core HTTP & UI dependencies
-gradio==4.44.1
-requests
-jsonschema>=4.0.0
-# Tokenizers & LLM
-tiktoken
-openai
+pandas==1.1.5
+pydantic<2.0
+gradio==3.39.0
+requests>=2.28