examples/docqa/doc-chat-2.py - main - githubmirror/langroid

examples/docqa/doc-chat-2.py
149 строк5 KB
Prasad Chalasani
Add TaskTool dynamic sub-agent spawn example + ruff auto-fix for examples (#876)
28 июн 2025, 00:02
28 июн 2025, 00:027169fc4
"""
2-agent doc-chat:
WriterAgent is in charge of answering user's question.
Breaks it down into smaller questions (if needed) to send to DocAgent,
who has access to the docs via a vector-db.

python3 examples/docqa/doc-chat-2.py
"""

import os

from fire import Fire
from rich import print
from rich.prompt import Prompt

import langroid as lr
import langroid.language_models as lm
from langroid.agent.chat_agent import ChatAgent, ChatAgentConfig
from langroid.agent.special.doc_chat_agent import (
    DocChatAgent,
    DocChatAgentConfig,
)
from langroid.agent.task import Task
from langroid.agent.tools.recipient_tool import RecipientTool
from langroid.mytypes import Entity
from langroid.parsing.parser import ParsingConfig, PdfParsingConfig, Splitter
from langroid.utils.configuration import Settings, set_global
from langroid.utils.constants import NO_ANSWER

os.environ["TOKENIZERS_PARALLELISM"] = "false"


def main(
    debug: bool = False,
    nocache: bool = False,
    model: str = lm.OpenAIChatModel.GPT4o,
) -> None:
    llm_config = lm.OpenAIGPTConfig(chat_model=model)
    config = DocChatAgentConfig(
        llm=llm_config,
        n_query_rephrases=0,
        hypothetical_answer=False,
        full_citations=False,
        assistant_mode=True,
        n_neighbor_chunks=2,
        n_similar_chunks=5,
        n_relevant_chunks=5,
        parsing=ParsingConfig(  # modify as needed
            splitter=Splitter.TOKENS,
            chunk_size=100,  # aim for this many tokens per chunk
            n_neighbor_ids=5,
            overlap=20,  # overlap between chunks
            max_chunks=10_000,
            # aim to have at least this many chars per chunk when
            # truncating due to punctuation
            min_chunk_chars=200,
            discard_chunk_chars=5,  # discard chunks with fewer than this many chars
            # NOTE: PDF parsing is extremely challenging, each library has its own
            # strengths and weaknesses. Try one that works for your use case.
            pdf=PdfParsingConfig(
                # alternatives: "unstructured", "docling", "fitz"
                library="pymupdf4llm",
            ),
        ),
    )

    embed_cfg = lr.embedding_models.OpenAIEmbeddingsConfig()

    config.vecdb = lr.vector_store.QdrantDBConfig(
        cloud=False,
        storage_path=".qdrant/doc-chat",
        embedding=embed_cfg,
    )

    set_global(
        Settings(
            debug=debug,
            cache=not nocache,
        )
    )

    doc_agent = DocChatAgent(config)
    print("[blue]Welcome to the document chatbot!")
    doc_agent.user_docs_ingest_dialog()
    print("[cyan]Enter x or q to quit, or ? for evidence")
    doc_task = Task(
        doc_agent,
        interactive=False,
        name="DocAgent",
        done_if_no_response=[Entity.LLM],  # done if null response from LLM
        done_if_response=[Entity.LLM],  # done if non-null response from LLM
    )

    writer_agent = ChatAgent(
        ChatAgentConfig(
            name="WriterAgent",
            llm=llm_config,
            vecdb=None,
        )
    )
    writer_agent.enable_message(RecipientTool)
    writer_task = Task(
        writer_agent,
        name="WriterAgent",
        interactive=False,
        system_message=f"""
        You are tenacious, creative and resourceful when given a question to 
        find an answer for. You will receive questions from a user, which you will 
        try to answer ONLY based on content from certain documents (not from your 
        general knowledge). However you do NOT have access to the documents. 
        You will be assisted by DocAgent, who DOES have access to the documents.
        
        Here are the rules:
        (a) when the question is complex or has multiple parts, break it into small 
         parts and/or steps and send them to DocAgent
        (b) if DocAgent says {NO_ANSWER} or gives no answer, try asking in other ways.
        (c) Once you collect all parts of the answer, say "DONE" 
            and show me the consolidated final answer. 
        (d) DocAgent has no memory of previous dialog, so you must ensure your 
            questions are stand-alone questions that don't refer to entities mentioned 
            earlier in the dialog.
        (e) if DocAgent is unable to answer after your best efforts, you can say
            {NO_ANSWER} and move on to the next question.
        (f) answers should be based ONLY on the documents, NOT on your prior knowledge.
        (g) be direct and concise, do not waste words being polite.
        (h) if you need more info from the user, before asking DocAgent, you should 
        address questions to the "User" (not to DocAgent) to get further 
        clarifications or information. 
        (i) Always ask questions ONE BY ONE (to either User or DocAgent), NEVER 
            send Multiple questions in one message.
        (j) Use bullet-point format when presenting multiple pieces of info.
        (k) When DocAgent responds without citing a SOURCE and EXTRACT(S), you should
            send your question again to DocChat, reminding it to cite the source and
            extract(s).
        
        Start by asking the user what they want to know.
        """,
    )
    writer_task.add_sub_task(doc_task)

    while True:
        query = Prompt.ask("[blue]How can I help?")
        if query in ["x", "q"]:
            break
        writer_task.run(query)


if __name__ == "__main__":
    Fire(main)