Day 5: Lets Build a Chat App with Bedrock

turn our Knowledge Base into a chat app with Python and Streamlit. Follow-up questions, source references, also see new releasees

Shaan

Published Oct 13, 2026

Last modified Oct 13, 2026

lets we build a small chat screen with:

A text box and chat history
Follow-up questions that continue the conversation
Source files you can check under every answer

It's a small Python app. No web development experience needed.

Who does what?

Here's what happens when you type a question:

Streamlit shows the chat box and sends your question to Bedrock.
The Knowledge Base searches your notes and finds the most relevant chunks.
The AI model reads those chunks and writes the answer.
The answer comes back to Streamlit along with the source files it used.

Piece	Its job
Amazon Bedrock	Provides the AI model and the Knowledge Base API
Knowledge Base	Finds the relevant parts of your notes
Streamlit	Draws the chat page
Conversation session	Lets follow-up questions build on earlier ones

The app runs on your laptop, with your AWS credentials, and calls the Knowledge Base

Which Knowledge Base? This article uses the classic Knowledge Base from Day 4 , where we chose the embedding model and vector store. AWS also launched a Managed Knowledge Base in 2026, covered near the end.

Before you start

A working Knowledge Base from Day 4 with a finished sync
Python and AWS credentials set up from Day 1
Your Knowledge Base ID and a model ARN or inference profile ARN that worked on Day 4

Step 1: Install Streamlit

Streamlit turns a Python file into a web page.

1
pip install --upgrade streamlit boto3

Step 2: Write the app

Create a file called app.py:

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
import boto3
import streamlit as st
from botocore.exceptions import ClientError

REGION = "YOUR_AWS_REGION"
KB_ID = "YOUR_KB_ID"
MODEL_ARN = "YOUR_MODEL_OR_INFERENCE_PROFILE_ARN"

client = boto3.client("bedrock-agent-runtime", region_name=REGION)

st.set_page_config(page_title="Chat with my notes")
st.title("Chat with my class notes")

if "messages" not in st.session_state:
    st.session_state.messages = []
    st.session_state.session_id = None

# Reset button: runs BEFORE the history is drawn
if st.button("Start a new chat"):
    st.session_state.messages = []
    st.session_state.session_id = None

def show_message(role, content, sources=None):
    with st.chat_message(role):
        st.markdown(content)
        if sources:
            with st.expander("View sources"):
                for s in sources:
                    st.write(s)

for m in st.session_state.messages:
    show_message(m["role"], m["content"], m.get("sources"))

question = st.chat_input("Ask something from your notes")

if question:
    st.session_state.messages.append({"role": "user", "content": question})
    show_message("user", question)

    params = {
        "input": {"text": question},
        "retrieveAndGenerateConfiguration": {
            "type": "KNOWLEDGE_BASE",
            "knowledgeBaseConfiguration": {
                "knowledgeBaseId": KB_ID,
                "modelArn": MODEL_ARN,
            },
        },
    }
    if st.session_state.session_id:
        params["sessionId"] = st.session_state.session_id

    try:
        with st.spinner("Searching your notes..."):
            response = client.retrieve_and_generate(**params)
        st.session_state.session_id = response["sessionId"]
        answer = response["output"]["text"]
        sources = sorted({
            ref["location"]["s3Location"]["uri"]
            for c in response.get("citations", [])
            for ref in c.get("retrievedReferences", [])
            if "s3Location" in ref.get("location", {})
        })
    except ClientError as e:
        answer = f"Something went wrong: {e.response['Error']['Message']}"
        sources = []

    show_message("assistant", answer, sources)
    st.session_state.messages.append(
        {"role": "assistant", "content": answer, "sources": sources}
    )

Replace the three placeholders at the top with your own values.

Step 3: Run it

1
streamlit run app.py

Your browser opens a chat page. Ask a question from your notes.

Two kinds of "memory"

Our app remembers things in two different places, and they do different jobs.

1. Streamlit remembers what to show. Every time you send a message, Streamlit reruns the whole script. Without st.session_state, the chat on screen would disappear after each question. This memory lives in your browser session. If you refresh the page, it starts fresh.

2. Bedrock remembers what the conversation is about. Each response includes a sessionId. When we send it back with the next question, Bedrock can understand follow-ups like "explain that in simpler words." It continues the conversation. It isn't a notebook of everything you've ever asked.

Memory	What it does	Limit
`st.session_state`	Keeps the chat messages on screen	A page refresh starts a new session
`sessionId` (Bedrock)	Gives the model context for follow-up	Only continues the conversation.

Try this

Ask: "What is a deadlock?"
Then ask: "Explain that in simpler words."
Click Start a new chat. This clears the messages on screen and the Bedrock sessionId, so the next question starts a brand-new conversation.
Ask "Explain that in simpler words." again.

If step 2 makes sense and step 4 doesn't, the session memory is working.

Experiment: is your AI actually using your notes?

A working chat is exciting. But I wanted to know: did the answer come from my notes, or did the model already know it?

Ask a question your notes answer clearly. Open View sources and check the file.
Ask about a topic that isn't in your notes. Compare the answer and the sources.
Create a small test file called test-fact.txt containing only this line.

If the app finds your made-up fact, retrieval is working.

My takeaway: an answer can sound confident without being supported by your documents. A source helps you check, but it doesn't prove every claim is right.

My result: The app answered my questions and showed the source files, making it easier to check where the information came from.

What the errors mean

What you see	What it means	What to try
`AccessDeniedException`	A permissions problem	Check IAM and model access in your console
`ValidationException` about the model	A model setup problem	Copy the exact model or inference profile ARN
No error, but "no information" answers	Retrieval found nothing useful	Check the sync status and whether the answer is in your notes

What's new in 2026 (Managed Knowledge Base)

AWS launched Amazon Bedrock Managed Knowledge Base, a fully managed option. These updates are for the managed version. Our tutorial uses the classic one.

June 17, 2026: generally available , with native connectors including Amazon S3
September 4, 2026: automatic sync scheduling (daily, weekly, or monthly)
September 23, 2026: Salesforce and Zendesk connectors

Classic vs Managed: they aren't interchangeable. Managed uses AWS-managed storage (no vector store choice), and its retrieval calls use a different configuration. Check Region availability and the current docs before switching. I kept this series on the classic version so you can see each building block.

Keep it safe and affordable

No access keys in your code. Use aws configure.
This app is for local learning. Sharing it with others would need hosting, login protection, and care for your AWS resources.
Costs depend on the model you use, the retrievals, and the services behind your Knowledge Base. Check the current AWS pricing and set a budget alert in AWS Budgets before experimenting.

Summary

A chat app is a small Streamlit page that calls your Knowledge Base
Streamlit session state and Bedrock's sessionId are different kinds of memory
Always check the sources, and test with facts you control
Classic and Managed Knowledge Bases are different options

Next: Day 6

Our bot will answer almost anything. Tomorrow we add Guardrails: block unsafe or off-topic questions and hide personal information.

What would you add to this chat app? Tell me below.

# amazon-bedrock # generative-ai # llm # python # amazon-bedrock-agents

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.

Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article

Builder Center

Day 5: Lets Build a Chat App with Bedrock

Shaan

Who does what?

Before you start

Step 1: Install Streamlit

Step 2: Write the app

Step 3: Run it

Two kinds of "memory"

Try this

Experiment: is your AI actually using your notes?

What the errors mean

What's new in 2026 (Managed Knowledge Base)

Keep it safe and affordable

Summary

Next: Day 6