setup

.gitignore

@@ -0,0 +1,3 @@
+/hfagent/
+/.vscode/
+.env

connectors.py

@@ -0,0 +1,7 @@
+import pandas as pd, requests
+
+def fetch_questions(url: str) -> pd.DataFrame:
+    response = requests.get(url)
+    response.raise_for_status()
+    question_data = response.json()
+    return pd.DataFrame(question_data)

local.ipynb

@@ -0,0 +1,332 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "dfd51b3d",
+   "metadata": {},
+   "source": [
+    "# Inizializzazione e Fetch Domande"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "660cfadd",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>task_id</th>\n",
+       "      <th>question</th>\n",
+       "      <th>Level</th>\n",
+       "      <th>file_name</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>count</th>\n",
+       "      <td>20</td>\n",
+       "      <td>20</td>\n",
+       "      <td>20</td>\n",
+       "      <td>20</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>unique</th>\n",
+       "      <td>20</td>\n",
+       "      <td>20</td>\n",
+       "      <td>1</td>\n",
+       "      <td>6</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>top</th>\n",
+       "      <td>8e867cd7-cff9-4e6c-867a-ff5ddc2550be</td>\n",
+       "      <td>How many studio albums were published by Merce...</td>\n",
+       "      <td>1</td>\n",
+       "      <td></td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>freq</th>\n",
+       "      <td>1</td>\n",
+       "      <td>1</td>\n",
+       "      <td>20</td>\n",
+       "      <td>15</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>"
+      ],
+      "text/plain": [
+       "                                     task_id  \\\n",
+       "count                                     20   \n",
+       "unique                                    20   \n",
+       "top     8e867cd7-cff9-4e6c-867a-ff5ddc2550be   \n",
+       "freq                                       1   \n",
+       "\n",
+       "                                                 question Level file_name  \n",
+       "count                                                  20    20        20  \n",
+       "unique                                                 20     1         6  \n",
+       "top     How many studio albums were published by Merce...     1            \n",
+       "freq                                                    1    20        15  "
+      ]
+     },
+     "execution_count": 1,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import os\n",
+    "from connectors import fetch_questions\n",
+    "\n",
+    "DEFAULT_API_URL = \"https://agents-course-unit4-scoring.hf.space\"\n",
+    "QUESTIONS_URL = f\"{DEFAULT_API_URL}/questions\"\n",
+    "SUBMIT_URL = f\"{DEFAULT_API_URL}/submit\"\n",
+    "HFTOKEN=os.getenv(\"HF_TOKEN\")\n",
+    "\n",
+    "questions = fetch_questions(QUESTIONS_URL)\n",
+    "questions.describe()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "41b1c193",
+   "metadata": {},
+   "source": [
+    "# Agent"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "02fe107c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "c:\\Users\\dgl17\\Desktop\\Lavoro\\Personal\\Final_Assignment_Template\\hfagent\\Lib\\site-packages\\tqdm\\auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
+      "  from .autonotebook import tqdm as notebook_tqdm\n"
+     ]
+    }
+   ],
+   "source": [
+    "from langgraph.graph import StateGraph, START, END\n",
+    "from langchain_huggingface import ChatHuggingFace, HuggingFacePipeline\n",
+    "from typing import TypedDict, Any\n",
+    "from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer\n",
+    "import torch\n",
+    "import random as rng"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4cb1b488",
+   "metadata": {},
+   "source": [
+    "## Schemas"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "01b8a368",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class ActionsSchema(TypedDict):\n",
+    "  name: str\n",
+    "  arguments: dict[str, Any] # name - value\n",
+    "  observation: Any | None\n",
+    "\n",
+    "class AgentSchema(TypedDict):\n",
+    "  query: str\n",
+    "  answer: str\n",
+    "  actions: list[ActionsSchema]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c8687b82",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Fetching 3 files:   0%|          | 0/3 [00:00<?, ?it/s]"
+     ]
+    }
+   ],
+   "source": [
+    "class RootAgent():\n",
+    "  __MODELID = \"mistralai/Mistral-7B-Instruct-v0.2\"\n",
+    "  \n",
+    "  def __init__(self):\n",
+    "    sg = StateGraph(AgentSchema)\n",
+    "    tokenizer = AutoTokenizer.from_pretrained(self.__MODELID)\n",
+    "    model = AutoModelForCausalLM.from_pretrained(self.__MODELID, dtype=torch.bfloat16, device_map=\"auto\")\n",
+    "    pipe = pipeline(\"text-classification\", model=model, tokenizer=tokenizer, temperature=0.3, do_sample=True, max_new_tokens=512)\n",
+    "    self.brain = ChatHuggingFace(llm=HuggingFacePipeline(pipeline=pipe))\n",
+    "    sg.add_node(\"plan\", self.__plan)\n",
+    "    sg.add_edge(START, \"plan\")\n",
+    "    sg.add_edge(\"plan\", END)\n",
+    "    self.agentflow = sg.compile()\n",
+    "    return\n",
+    "  \n",
+    "  def __call__(self, query: str):\n",
+    "    return self.agentflow.invoke({\"actions\": [], \"query\": query, \"answer\": \"\"})\n",
+    "  \n",
+    "  ## nodes ##\n",
+    "  def __plan(self, state: AgentSchema) -> AgentSchema:\n",
+    "    stageprompt = f\"\"\"\n",
+    "      The user asked the following question: {state.get(\"query\", \"\")}.\n",
+    "      In past iterations you observed the following results: {state.get(\"actions\", [])}\n",
+    "\n",
+    "      You are the planning manager entitled to develop an efficient plan to obtain the answer, for this reason you must follow these rules:\n",
+    "      1. IF you can answer directly without using any tools, given the information you have, generate the [Answer Structured JSON](#Answer-Structured-JSON) format\n",
+    "      2. OTHERWISE generate a plan consisting of tool calls for which you have all the data to pass as arguments following the [Plan Structured JSON](#Plan-Structured-JSON) format\n",
+    "\n",
+    "      !IMPORTANT: ALWAYS use one of the two JSON format for your answer!\n",
+    "      \n",
+    "      #Answer Structured JSON\n",
+    "      ```json\n",
+    "      {\"{'Answer': '<answer1>,<answer2>,...'}\"}\n",
+    "      ```\n",
+    "      !IMPORTANT: Only provide the comma separated answers that are directly asked by the user, without any formatting or measurement unit!\n",
+    "      ## Examples\n",
+    "      1. user: 'what is the capital of Italy?' --> {\"{'Answer': 'Rome'}\"}\n",
+    "      2. user: 'What is the nearest city to Firenze? How many kilometers? --> past iterations: [{\"{'name': 'web_search', 'arguments': {'query': 'Distance between Firenze and Pisa in Kilometers'}, 'observation': '150 KM'}\"}] --> {\"{'Answer': 'Pisa,150'}\"}\n",
+    "\n",
+    "      #Plan Structured JSON\n",
+    "      ```json\n",
+    "      {\"\"\"{'Plan': [\n",
+    "        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1, '<argument_name_2>': argument_value_2, ...}},\n",
+    "        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1}},\n",
+    "        ...\n",
+    "      ]}\"\"\"}\n",
+    "      ```\n",
+    "      !IMPORTANT: Those will be used as function calls so be extremely precise on argument<-->value mappings, moreover to better understand\n",
+    "      the task, always place a Reasoning string before the plan JSON\n",
+    "      ## Examples\n",
+    "      1. user: 'What is the distance between Rome and Paris?' --> Reasoning: To get the distance I need to search the web, since it is a very\n",
+    "      specific question and I can choose between web_search and wiki_search, I think it is better to search in the web with query='distance Rome to Paris in Kilometers' --> {\"\"\"{\n",
+    "        'Plan': [\n",
+    "          {'name': '<web_search>', 'arguments': {'query': 'distance Rome to Paris'}},\n",
+    "          ...\n",
+    "        ]\n",
+    "      }\"\"\"}\n",
+    "    \"\"\"\n",
+    "    response = self.brain.invoke(stageprompt)\n",
+    "    return {\"query\": state.get(\"query\"), \"answer\": response, \"actions\": []}\n",
+    "\n",
+    "agent = RootAgent()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "64fb34d8",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The user asked the following question: Who are the pitchers with the number before and after Taishō Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters..\n",
+      "      In past iterations you observed the following results: []\n",
+      "\n",
+      "      You are the planning manager entitled to develop an efficient plan to obtain the answer, for this reason you must follow these rules:\n",
+      "      1. IF you can answer directly without using any tools, given the information you have, generate the [Answer Structured JSON](#Answer-Structured-JSON) format\n",
+      "      2. OTHERWISE generate a plan consisting of tool calls for which you have all the data to pass as arguments following the [Plan Structured JSON](#Plan-Structured-JSON) format\n",
+      "\n",
+      "      !IMPORTANT: ALWAYS use one of the two JSON format for your answer!\n",
+      "\n",
+      "      #Answer Structured JSON\n",
+      "      ```json\n",
+      "      {'Answer': '<answer1>,<answer2>,...'}\n",
+      "      ```\n",
+      "      !IMPORTANT: Only provide the comma separated answers that are directly asked by the user, without any formatting or measurement unit!\n",
+      "      ## Examples\n",
+      "      1. user: 'what is the capital of Italy?' --> {'Answer': 'Rome'}\n",
+      "      2. user: 'What is the nearest city to Firenze? How many kilometers? --> past iterations: [{'name': 'web_search', 'arguments': {'query': 'Distance between Firenze and Pisa in Kilometers'}, 'observation': '150 KM'}] --> {'Answer': 'Pisa,150'}\n",
+      "\n",
+      "      #Plan Structured JSON\n",
+      "      ```json\n",
+      "      {'Plan': [\n",
+      "        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1, '<argument_name_2>': argument_value_2, ...}},\n",
+      "        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1}},\n",
+      "        ...\n",
+      "      ]}\n",
+      "      ```\n",
+      "      !IMPORTANT: Those will be used as function calls so be extremely precise on argument<-->value mappings, moreover to better understand\n",
+      "      the task, always place a Reasoning string before the plan JSON\n",
+      "      ## Examples\n",
+      "      1. user: 'What is the distance between Rome and Paris?' --> Reasoning: To get the distance I need to search the web, since it is a very\n",
+      "      specific question and I can choose between web_search and wiki_search, I think it is better to search in the web with query='distance Rome to Paris in Kilometers' --> {\n",
+      "        'Plan': [\n",
+      "          {'name': '<web_search>', 'arguments': {'query': 'distance Rome to Paris'}},\n",
+      "          ...\n",
+      "        ]\n",
+      "      }\n",
+      "{'query': \"Who are the pitchers with the number before and after Taishō Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters.\", 'answer': \"The user asked the following question: Who are the pitchers with the number before and after Taishō Tamai's number as of July 2023? Give them to me in the form Pitcher Before, Pitcher After, use their last names only, in Roman characters..\\n      In past iterations you observed the following results: []\\n\\n      You are the planning manager entitled to develop an efficient plan to obtain the answer, for this reason you must follow these rules:\\n      1. IF you can answer directly without using any tools, given the information you have, generate the [Answer Structured JSON](#Answer-Structured-JSON) format\\n      2. OTHERWISE generate a plan consisting of tool calls for which you have all the data to pass as arguments following the [Plan Structured JSON](#Plan-Structured-JSON) format\\n\\n      !IMPORTANT: ALWAYS use one of the two JSON format for your answer!\\n\\n      #Answer Structured JSON\\n      ```json\\n      {'Answer': '<answer1>,<answer2>,...'}\\n      ```\\n      !IMPORTANT: Only provide the comma separated answers that are directly asked by the user, without any formatting or measurement unit!\\n      ## Examples\\n      1. user: 'what is the capital of Italy?' --> {'Answer': 'Rome'}\\n      2. user: 'What is the nearest city to Firenze? How many kilometers? --> past iterations: [{'name': 'web_search', 'arguments': {'query': 'Distance between Firenze and Pisa in Kilometers'}, 'observation': '150 KM'}] --> {'Answer': 'Pisa,150'}\\n\\n      #Plan Structured JSON\\n      ```json\\n      {'Plan': [\\n        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1, '<argument_name_2>': argument_value_2, ...}},\\n        {'name': '<tool_name>', 'arguments': {'<argument_name_1>': argument_value_1}},\\n        ...\\n      ]}\\n      ```\\n      !IMPORTANT: Those will be used as function calls so be extremely precise on argument<-->value mappings, moreover to better understand\\n      the task, always place a Reasoning string before the plan JSON\\n      ## Examples\\n      1. user: 'What is the distance between Rome and Paris?' --> Reasoning: To get the distance I need to search the web, since it is a very\\n      specific question and I can choose between web_search and wiki_search, I think it is better to search in the web with query='distance Rome to Paris in Kilometers' --> {\\n        'Plan': [\\n          {'name': '<web_search>', 'arguments': {'query': 'distance Rome to Paris'}},\\n          ...\\n        ]\\n      }\", 'actions': []}\n"
+     ]
+    },
+    {
+     "ename": "",
+     "evalue": "",
+     "output_type": "error",
+     "traceback": [
+      "\u001b[1;31mnotebook controller is DISPOSED. \n",
+      "\u001b[1;31mView Jupyter <a href='command:jupyter.viewOutput'>log</a> for further details."
+     ]
+    }
+   ],
+   "source": [
+    "query = questions[\"question\"].iloc[rng.randint(0, len(questions[\"question\"]))]\n",
+    "print(agent(query))"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.13"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}