File size: 8,652 Bytes
1b8aef5
 
 
 
 
 
 
 
 
 
 
b066853
1b8aef5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b066853
 
 
 
 
 
 
 
 
 
 
 
 
1b8aef5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
import re
from typing import Any, Literal

from langchain_community.tools import DuckDuckGoSearchResults
from langchain_core.messages import SystemMessage, AnyMessage
from langchain_core.runnables import Runnable
from langchain_core.tools import BaseTool
from langchain_ollama import ChatOllama
from langgraph.constants import START, END
from langgraph.graph import MessagesState, StateGraph
from langgraph.graph.graph import CompiledGraph
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
from langgraph.prebuilt import ToolNode
from pydantic import BaseModel

from tools import (
    get_excel_table_content,
    get_youtube_video_transcript,
    reverse_string,
    transcribe_audio_file,
    web_page_info_retriever,
    youtube_video_to_frame_captions, sum_list, execute_python_script,
)


class AgentFactory:
    """
    A factory for the agent. It is assumed that an Ollama server is running
    on the machine where the factory is used.
    """

    __system_prompt: str = (
        "You have to answer to test questions and you need to score high.\n"
        "Sometimes auxiliary files may be attached to the question so the\n"
        "question itself is presented as a JSON string with the following\n"
        "fields:\n"
        "1. task_id: unique hash identifier of the question.\n"
        "2. question: the text of the question.\n"
        "3. Level: a number with the question difficulty level. You can ignore "
        "this field.\n"
        "4. file_name: the name of the file needed to answer the question. "
        "This is empty if the question does not refer to any file. "
        "IMPORTANT: The text of the question may mention a file name that is "
        "different from what is reported into the \"file_name\" JSON field. "
        "YOU HAVE TO IGNORE THE FILE NAME MENTIONED INTO \"question\" AND "
        "YOU MUST USE THE FILE NAME PROVIDED INTO THE \"file_name\" FIELD.\n"
        "\n"
        "Depending on the question, the\n"
        "format of your answer is a number OR as few words as possible OR a\n"
        "comma separated list of numbers and/or strings. If you are asked for\n"
        "a number, don't use comma to write your number neither use units\n"
        "such as $ or percent sign unless specified otherwise. If you are\n"
        "asked for a string, don't use articles, neither abbreviations (e.g.\n"
        "for cities), and write the digits in plain text unless specified\n"
        "otherwise. If you are asked for a comma separated list, apply the\n"
        "above rules depending of whether the element to be put in the list\n"
        "is a number or a string.\n"
        "When you have to perform a sum, DON'T try to do that yourself.\n"
        "Exploit the tool that is able to sum list of numbers. If you have\n"
        "to sum the results of previous sums, use again the same tool\n"
        "recursively. NEVER do the sums yourself.\n"
        "Achieve the solution by dividing your reasoning in steps, and\n"
        "provide an explanation for each step.\n"
        "You are advised to cycle between reasoning and tool calling also\n"
        "multiple times. Provide an answer only when you are sure you don't\n"
        "have to call any tool again. Provide the answer between\n"
        "<ANSWER> and </ANSWER> tags. I stress that the final answer must\n"
        "follow the rules explained above.\n"
    )

    __llm: Runnable
    __tools: list[BaseTool]

    def __init__(
        self,
        model: str = "qwen2.5-coder:32b",
        # model: str = "mistral-small3.1",
        # model: str = "phi4-mini",
        temperature: float = 0.0,
        num_ctx: int = 8192
    ) -> None:
        """
        Constructor.

        Args:
            model:          The name of the Ollama model to use.
            temperature:    Temperature parameter.
            num_ctx:        Size of the context window used to generate the
                            next token.
        """
        search_tool = DuckDuckGoSearchResults(
            description=(
                "A wrapper around Duck Duck Go Search. Useful for when you "
                "need to answer questions about information you can find on "
                "the web. Input should be a search query. It is advisable to "
                "use this tool to retrieve web page URLs and use another tool "
                "to analyze the pages. If the web source is suggested by the "
                "user query, prefer retrieving information from that source. "
                "For example, the query may suggest to search on Wikipedia or "
                "Medium. In those cases, prepend the query with "
                "'site: <name of the source>'. For example: "
                "'site: wikipedia.org'"
            ),
            output_format="list"
        )
        search_tool.with_retry()
        self.__tools = [
            execute_python_script,
            get_excel_table_content,
            get_youtube_video_transcript,
            reverse_string,
            search_tool,
            sum_list,
            transcribe_audio_file,
            web_page_info_retriever,
            youtube_video_to_frame_captions
        ]
        self.__llm = ChatOllama(
            model=model,
            temperature=temperature,
            num_ctx=num_ctx
        ).bind_tools(tools=self.__tools)

        # llm_endpoint = HuggingFaceEndpoint(
        #     repo_id="Qwen/Qwen2.5-72B-Instruct",
        #     task="text-generation",
        #     max_new_tokens=num_ctx,
        #     do_sample=False,
        #     repetition_penalty=1.03,
        #     temperature=temperature,
        # )
        #
        # self.__llm = (
        #     ChatHuggingFace(llm=llm_endpoint)
        #         .bind_tools(tools=self.__tools)
        # )

    def __run_llm(self, state: MessagesState) -> dict[str, Any]:
        answer = self.__llm.invoke(state["messages"])
        # Remove thinking pattern if present
        pattern = r'\n*<think>.*?</think>\n*'
        answer.content = re.sub(
            pattern, "", answer.content, flags=re.DOTALL
        )
        return {"messages": [answer]}

    @staticmethod
    def __extract_last_message(
        state: list[AnyMessage] | dict[str, Any] | BaseModel,
        messages_key: str
    ) -> str:
        if isinstance(state, list):
            last_message = state[-1]
        elif isinstance(state, dict) and (messages := state.get(messages_key, [])):
            last_message = messages[-1]
        elif messages := getattr(state, messages_key, []):
            last_message = messages[-1]
        else:
            raise ValueError(f"No messages found in input state to tool_edge: {state}")
        return last_message

    def __route_from_llm(
        self,
        state: list[AnyMessage] | dict[str, Any] | BaseModel,
        messages_key: str = "messages",
    ) -> Literal["tools", "extract_final_answer"]:
        ai_message = self.__extract_last_message(state, messages_key)
        if hasattr(ai_message, "tool_calls") and len(ai_message.tool_calls) > 0:
            return "tools"
        return "extract_final_answer"

    @staticmethod
    def __extract_final_answer(state: MessagesState) -> dict[str, Any]:
        last_message = state["messages"][-1].content
        pattern = r"<ANSWER>(?P<answer>.*?)</ANSWER>"
        m = re.search(pattern, last_message, flags=re.DOTALL)
        answer = m.group("answer").strip() if m else ""
        return {"messages": [answer]}

    @property
    def system_prompt(self) -> SystemMessage:
        """
        Returns:
            The system prompt to use with the agent.
        """
        return SystemMessage(content=self.__system_prompt)

    def get(self) -> CompiledGraph:
        """
        Factory method.

        Returns:
            The instance of the agent.
        """
        graph_builder = StateGraph(MessagesState)

        graph_builder.add_node("LLM", self.__run_llm)
        graph_builder.add_node("tools", ToolNode(tools=self.__tools))
        graph_builder.add_node(
            "extract_final_answer",
            self.__extract_final_answer
        )

        graph_builder.add_edge(start_key=START, end_key="LLM")
        graph_builder.add_conditional_edges(
            source="LLM",
            path=self.__route_from_llm,
            path_map={
                "tools": "tools",
                "extract_final_answer": "extract_final_answer"
            }
        )
        graph_builder.add_edge(start_key="tools", end_key="LLM")
        graph_builder.add_edge(start_key="extract_final_answer", end_key=END)

        return graph_builder.compile()