Added voting logic (#9)

* Install llm-core

* Use core in RMQ

* Use config.py as core duplicate

* Use core in model

* Updated diana config

* Updated readme

* Fix embeddings output type

* Fix model variable duplicate

* Added num_parallel_processes parameter

* Fix Dict type def

* Init all abstract methods

* Fix _score call

* chatgpt -> chat_gpt

* Use pass instead of None

README.md

@@ -25,15 +25,16 @@ MQ:
   port: <MQ Port>
   server: <MQ Hostname or IP>
   users:
-    mq-chatgpt-api:
+    neon_llm_chat_gpt:
       password: <neon_chatgpt user's password>
       user: neon_chatgpt
-ChatGPT:
+LLM_CHAT_GPT:
   key: ""
   model: "gpt-3.5-turbo"
   role: "You are trying to give a short answer in less than 40 words."
   context_depth: 3
   max_tokens: 100
+  num_parallel_processes: 2
 ```
 
 For example, if your configuration resides in `~/.config`:

docker_overlay/etc/neon/diana.yaml

@@ -14,8 +14,9 @@ MQ:
     mq_handler:
       user: neon_api_utils
       password: Klatchat2021
-ChatGPT:
+LLM_CHAT_GPT:
   model: "gpt-3.5-turbo"
   role: "You are trying to give a short answer in less than 40 words."
   context_depth: 3
-  max_tokens: 100
+  max_tokens: 100
+  num_parallel_processes: 2

neon_llm_chatgpt/chatgpt.py

@@ -25,39 +25,132 @@
 # SOFTWARE,  EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 import openai
+from openai.embeddings_utils import get_embeddings, distances_from_embeddings
 
+from typing import List, Dict
+from neon_llm_core.llm import NeonLLM
+
+
+class ChatGPT(NeonLLM):
+
+    mq_to_llm_role = {
+        "user": "user",
+        "llm": "assistant"
+    }
 
-class ChatGPT:
     def __init__(self, config):
-        self.model = config["model"]
+        super().__init__(config)
+        self.model_name = config["model"]
         self.role = config["role"]
         self.context_depth = config["context_depth"]
         self.max_tokens = config["max_tokens"]
-        openai.api_key = config["key"]
+        self.api_key = config["key"]
+        self.num_parallel_processes = config["num_parallel_processes"]
+        self.warmup()
+
+    @property
+    def tokenizer(self) -> None:
+        return self._tokenizer
+
+    @property
+    def tokenizer_model_name(self) -> str:
+        return ""
+
+    @property
+    def model(self) -> openai:
+        if self._model is None:
+            openai.api_key = self.api_key
+            self._model = openai
+        return self._model
+
+    @property
+    def llm_model_name(self) -> str:
+        return self.model_name
 
-    @staticmethod
-    def convert_role(role):
-        if role == "user":
-            role_chatgpt = "user"
-        elif role == "llm":
-            role_chatgpt = "assistant"
-        return role_chatgpt
+    @property
+    def _system_prompt(self) -> str:
+        return self.role
 
-    def ask(self, message, chat_history):
+    def warmup(self):
+        self.model
+
+    def get_sorted_answer_indexes(self, question: str, answers: List[str]) -> List[int]:
+        """
+            Creates sorted list of answer indexes with respect to order provided in :param answers based on PPL score
+            Answers are sorted from best to worst
+            :param question: incoming question
+            :param answers: list of answers to rank
+            :returns list of indexes
+        """
+        if not answers:
+            return []
+        scores = self._score(prompt=question, targets=answers)
+        sorted_items = sorted(zip(range(len(answers)), scores), key=lambda x: x[1])
+        sorted_items_indexes = [x[0] for x in sorted_items]
+        return sorted_items_indexes
+
+    def _call_model(self, prompt: List[Dict[str, str]]) -> str:
+        """
+            Wrapper for ChatGPT Model generation logic
+            :param prompt: Input messages sequence
+            :returns: Output text sequence generated by model
+        """
+
+        response = openai.ChatCompletion.create(
+            model=self.llm_model_name,
+            messages=prompt,
+            temperature=0,
+            max_tokens=self.max_tokens,
+        )
+        text = response.choices[0].message['content']
+
+        return text
+
+    def _assemble_prompt(self, message: str, chat_history: List[List[str]]) -> List[Dict[str, str]]:
+        """
+            Assembles prompt engineering logic
+            Setup Guidance:
+            https://platform.openai.com/docs/guides/gpt/chat-completions-api
+
+            :param message: Incoming prompt
+            :param chat_history: History of preceding conversation
+            :returns: assembled prompt
+        """
         messages = [
-            {"role": "system", "content": self.role},
+            {"role": "system", "content": self._system_prompt},
         ]
         # Context N messages
         for role, content in chat_history[-self.context_depth:]:
             role_chatgpt = self.convert_role(role)
             messages.append({"role": role_chatgpt, "content": content})
         messages.append({"role": "user", "content": message})
-
-        response = openai.ChatCompletion.create(
-            model=self.model,
-            messages=messages,
-            temperature=0,
-            max_tokens=self.max_tokens,
-        )
-        bot_message = response.choices[0].message['content']
-        return bot_message
+        return messages
+
+    def _score(self, prompt: str, targets: List[str]) -> List[float]:
+        """
+            Calculates logarithmic probabilities for the list of provided text sequences
+            :param prompt: Input text sequence
+            :param targets: Output text sequences
+            :returns: List of calculated logarithmic probabilities per output text sequence
+        """
+
+        question_embeddings, answers_embeddings = self._embeddings(question=prompt, answers=targets)
+        scores_list = distances_from_embeddings(question_embeddings, answers_embeddings)
+        return scores_list
+
+    def _tokenize(self, prompt: str) -> None:
+        pass
+
+    def _embeddings(self, question: str, answers: List[str]) -> (List[float], List[List[float]]):
+        """
+            Computes embeddings for the list of provided answers
+            :param question: Question for LLM to response to
+            :param answers: List of provided answers
+            :returns ppl values for each answer
+        """
+        response = self.ask(question, [])
+        texts = [response] + answers
+        embeddings = get_embeddings(texts, engine="text-embedding-ada-002")
+        question_embeddings = embeddings[0]
+        answers_embeddings = embeddings[1:]
+        return question_embeddings, answers_embeddings

neon_llm_chatgpt/rmq.py

@@ -23,71 +23,33 @@
 # LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
 # NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 # SOFTWARE,  EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-
-import pika
-
-from neon_mq_connector.connector import MQConnector
-from neon_mq_connector.utils.network_utils import dict_to_b64
-from neon_mq_connector.utils.rabbit_utils import create_mq_callback
-from ovos_utils.log import LOG
+from neon_llm_core.rmq import NeonLLMMQConnector
 
 from neon_llm_chatgpt.chatgpt import ChatGPT
-from neon_llm_chatgpt.config import load_config
 
 
-class ChatgptMQ(MQConnector):
+class ChatgptMQ(NeonLLMMQConnector):
+    """
+        Module for processing MQ requests to ChatGPT
     """
-    Module for processing MQ requests from PyKlatchat to LibreTranslate"""
 
     def __init__(self):
-        config = load_config()
-        chatgpt_config = config.get("ChatGPT", None)
-        self.chatGPT = ChatGPT(chatgpt_config)
-
-        self.service_name = 'neon_llm_chatgpt'
-
-        mq_config = config.get("MQ", None)
-        super().__init__(config=mq_config, service_name=self.service_name)
-
-        self.vhost = "/llm"
-        self.queue = "chat_gpt_input"
-        self.register_consumer(name=self.service_name,
-                               vhost=self.vhost,
-                               queue=self.queue,
-                               callback=self.handle_request,
-                               on_error=self.default_error_handler,
-                               auto_ack=False)
-
-    @create_mq_callback(include_callback_props=('channel', 'method', 'body'))
-    def handle_request(self,
-                       channel: pika.channel.Channel,
-                       method: pika.spec.Basic.Return,
-                       body: dict):
-        """
-        Handles requests from MQ to ChatGPT received on queue
-        "request_chatgpt"
-
-        :param channel: MQ channel object (pika.channel.Channel)
-        :param method: MQ return method (pika.spec.Basic.Return)
-        :param body: request body (dict)
-        """
-        message_id = body["message_id"]
-        routing_key = body["routing_key"]
+        super().__init__()
+        self.warmup()
 
-        query = body["query"]
-        history = body["history"]
+    @property
+    def name(self):
+        return "chat_gpt"
 
-        response = self.chatGPT.ask(message=query, chat_history=history)
+    @property
+    def model(self):
+        if self._model is None:
+            self._model = ChatGPT(self.model_config)
+        return self._model
 
-        api_response = {
-            "message_id": message_id,
-            "response": response
-        }
+    def warmup(self):
+        self.model
 
-        channel.basic_publish(exchange='',
-                              routing_key=routing_key,
-                              body=dict_to_b64(api_response),
-                              properties=pika.BasicProperties(
-                                  expiration=str(1000)))
-        channel.basic_ack(method.delivery_tag)
-        LOG.info(f"Handled request: {message_id}")
+    @staticmethod
+    def compose_opinion_prompt(respondent_nick: str, question: str, answer: str) -> str:
+        return f'Why Answer "{answer}" to the Question "{question}" generated by Bot named "{respondent_nick}" is good?'

requirements/requirements.txt

@@ -1,4 +1,4 @@
+# model
 openai~=0.27
-neon-mq-connector~=0.7
-ovos-utils~=0.0.32
-ovos-config~=0.0.10
+# networking
+neon_llm_core==0.0.6