Knowledge Base for Amazon Bedrockの新機能をためす

雑な私をお許しください。

こちらが超詳しいやつです。

Streamlitにしただけです、すいません。

import boto3
import streamlit as st

client = boto3.client("bedrock-agent-runtime")

st.title("Chat with your document")
st.write("Knowledge Bases for Amazon Bedrock")

uploaded_file = st.file_uploader(
    "The supported file types are PDF, MD, TXT, DOCX, HTML, CSV, XLS, and XLSX",
    type=["PDF", "MD", "TXT", "DOCX", "HTML", "CSV", "XLS", "XLSX"],
)
if uploaded_file:
    uploaded_file_bytes = uploaded_file.getvalue()
    uploaded_file_name = uploaded_file.name
    uploaded_file_type = uploaded_file.type

if "messages" not in st.session_state:
    st.session_state.messages = []

for message in st.session_state.messages:
    with st.chat_message(message["role"]):
        st.markdown(message["content"])
        if "citations" in message:
            with st.expander("citations"):
                st.json(message["citations"])

if prompt := st.chat_input(
    "何でも聞いて下さい" if uploaded_file else "ファイルを選択してください",
    disabled=(not uploaded_file),
):
    st.session_state.messages.append({"role": "user", "content": prompt})
    with st.chat_message("user"):
        st.markdown(prompt)

    with st.chat_message("assistant"):

        params = {
            "input": {"text": prompt},
            "retrieveAndGenerateConfiguration": {
                "type": "EXTERNAL_SOURCES",
                "externalSourcesConfiguration": {
                    "modelArn": "anthropic.claude-3-sonnet-20240229-v1:0",
                    "sources": [
                        {
                            "sourceType": "BYTE_CONTENT",
                            "byteContent": {
                                "contentType": uploaded_file_type,
                                "data": uploaded_file_bytes,
                                "identifier": uploaded_file_name,
                            },
                        }
                    ],
                },
            },
        }

        if "sessionId" in st.session_state:
            params["sessionId"] = st.session_state["sessionId"]

        response = client.retrieve_and_generate(**params)

        st.markdown(response["output"]["text"])
        with st.expander("citations"):
            st.json(response["citations"])

        st.session_state["sessionId"] = response["sessionId"]
        st.session_state.messages.append(
            {
                "role": "assistant",
                "content": response["output"]["text"],
                "citations": response["citations"],
            }
        )

ファイルサイズは10MBまでなので、オプションを指定して起動

streamlit run app.py --server.maxUploadSize 10

我慢できずにやってしまった😆

・10MB以内でも文字数が多いとエラー
・チャンク分割されて処理されるので全体要約はうまくいかないかも
・待ち時間は結構ある pic.twitter.com/HrYLELJT8l

— moritalous | Kazuaki Morita (@moritalous) April 24, 2024