This documentation page is also available as an interactive notebook. You can launch the notebook in
Kaggle or Colab, or download it for use with an IDE or local Jupyter installation, by clicking one of the
above links.
Prerequisites
- Activate Bedrock in your AWS account.
- Request access to your desired models (e.g. Claude Sonnet 3.7,
Amazon Nova Pro)
- Optional - you may need to configure AWS CLI locally to authenticate
with your AWS account.
Important Notes
- Bedrock usage may incur costs based on your Bedrock plan.
- Be mindful of sensitive data and consider security measures when
integrating with external services.
First you’ll need to install required libraries and enter an Bedrock API
key.
%pip install -qU pixeltable boto3 duckduckgo-search
import os
import getpass
if 'AWS_ACCESS_KEY' not in os.environ:
os.environ['AWS_ACCESS_KEY'] = getpass.getpass('Enter your AWS Access Key:')
if 'AWS_SECRET_ACCESS_KEY' not in os.environ:
os.environ['AWS_SECRET_ACCESS_KEY'] = getpass.getpass('Enter your AWS Secret Access Key:')
import pixeltable as pxt
# Remove the `bedrock_demo` directory and its contents, if it exists
pxt.drop_dir('bedrock_demo', force=True)
pxt.create_dir('bedrock_demo')
Created directory ‘bedrock_demo’.
<pixeltable.catalog.dir.Dir at 0x1d4253ae930>
Basic Messages
Create a Table: In Pixeltable, create a table with columns to represent
your input data and the columns where you want to store the results from
Bedrock.
from pixeltable.functions import bedrock
# Create a table in Pixeltable and pick a model hosted on Bedrock with some parameters
t = pxt.create_table('bedrock_demo.chat', {'input': pxt.String})
msgs = [{'role': 'user', 'content': t.input}]
t.add_computed_column(output=bedrock.converse(
model_id="amazon.nova-pro-v1:0",
messages=[
{
'role': 'user',
'content': [
{
'text': t.input,
}
]
}
],
))
Created table `chat`.
Added 0 column values with 0 errors.
UpdateStatus(num_rows=0, num_computed_values=0, num_excs=0, updated_cols=[], cols_with_excs=[])
# Parse the response into a new column
t.add_computed_column(response=t.output.output.message.content[0].text)
Added 0 column values with 0 errors.
UpdateStatus(num_rows=0, num_computed_values=0, num_excs=0, updated_cols=[], cols_with_excs=[])
# Start a conversation
t.insert([{'input': 'What was the outcome of the 1904 US Presidential election?'}])
t.select(t.input, t.response).show()
Inserting rows into `chat`: 1 rows [00:00, 147.54 rows/s]
Inserted 1 row with 0 errors.
Now let’s create a more advanced example using Bedrock with tools for
news search and weather information.
import pixeltable as pxt
import pixeltable.functions as pxtf
from pixeltable.functions.bedrock import converse, invoke_tools
from duckduckgo_search import DDGS
# Initialize app structure
pxt.drop_dir("agents", force=True)
pxt.create_dir("agents")
Created directory ‘agents’.
<pixeltable.catalog.dir.Dir at 0x1d4252acdd0>
# Define tools
@pxt.udf
def search_news(keywords: str, max_results: int) -> str:
"""Search news using DuckDuckGo and return results."""
try:
with DDGS() as ddgs:
results = ddgs.news(
keywords=keywords,
region="wt-wt",
safesearch="off",
timelimit="m",
max_results=max_results,
)
formatted_results = []
for i, r in enumerate(results, 1):
formatted_results.append(
f"{i}. Title: {r['title']}\n"
f" Source: {r['source']}\n"
f" Published: {r['date']}\n"
f" Snippet: {r['body']}\n"
)
return "\n".join(formatted_results)
except Exception as e:
return f"Search failed: {str(e)}"
@pxt.udf
def get_weather(location: str) -> str:
"""Mock weather function - replace with actual API call."""
return f"Current weather in {location}: 72°F, Partly Cloudy"
# Register all tools
tools = pxt.tools(search_news, get_weather)
# Create base table
tool_agent = pxt.create_table(
"agents.tools",
{"prompt": pxt.String},
if_exists="ignore"
)
# Add tool selection and execution workflow
tool_agent.add_computed_column(
initial_response=converse(
model_id="amazon.nova-pro-v1:0",
messages=[
{
'role': 'user',
'content': [
{
'text': tool_agent.prompt,
}
]
}
],
tool_config=tools,
)
)
# Add tool execution
tool_agent.add_computed_column(
tool_output=invoke_tools(tools, tool_agent.initial_response)
)
Created table `tools`.
Added 0 column values with 0 errors.
Added 0 column values with 0 errors.
UpdateStatus(num_rows=0, num_computed_values=0, num_excs=0, updated_cols=[], cols_with_excs=[])
# Add response formatting
tool_agent.add_computed_column(
tool_response_prompt=pxtf.string.format(
"Orginal Prompt\n{0}: Tool Output\n{1}",
tool_agent.prompt,
tool_agent.tool_output
),
if_exists="ignore",
)
# Add final response generation
tool_agent.add_computed_column(
final_response=converse(
model_id="amazon.nova-pro-v1:0",
messages=[
{
'role': 'user',
'content': [
{
'text': tool_agent.tool_response_prompt,
}
]
}
]
)
)
tool_agent.add_computed_column(
answer=tool_agent.final_response.output.message.content[0].text
)
Added 0 column values with 0 errors.
Added 0 column values with 0 errors.
Added 0 column values with 0 errors.
UpdateStatus(num_rows=0, num_computed_values=0, num_excs=0, updated_cols=[], cols_with_excs=[])
# Example queries using different tools
queries = [
"What's the latest news about SpaceX?",
"What's the weather in San Francisco?",
]
# Use the agent
for query in queries:
tool_agent.insert([{'prompt': query}])
result = tool_agent.select(
tool_agent.prompt,
tool_agent.tool_output,
tool_agent.answer
).tail(1)
print(f"\nQuery: {query}")
print(f"Answer: {result['answer'][0]}")
# Display the full table
tool_agent.select(tool_agent.prompt, tool_agent.answer).show()
Inserting rows into `tools`: 1 rows [00:00, 168.89 rows/s]
Inserted 1 row with 0 errors.Query: What’s the latest news about SpaceX?
Answer: Here’s the latest news about SpaceX:1. Starbase City Proposal:
- Title: The home of Elon Musk’s SpaceX could become an official Texas city called Starbase
- Source: The Associated Press
- Published: 2025-04-30
- Snippet: An election is scheduled for Saturday to determine whether a small area of coastal South Texas, which is home to SpaceX, will become an official city named Starbase.2. 250th Starlink Satellite Mission:
- Title: SpaceX launches 250th Starlink satellite mission, lands rocket at sea (video, photos)
- Source: Space.com
- Published: 2025-04-28
- Snippet: SpaceX achieved a milestone by launching its 250th Starlink mission. A Falcon 9 rocket carrying 23 Starlink satellites, including 13 with direct-to-cell capability, lifted off from Cape Canaveral Space Force Station.3. Back-to-Back Starlink Launches:
- Title: SpaceX Falcon 9 rocket launches 1st of 2 planned Starlink launches in 2 days, lands booster at sea (video)
- Source: Space.com
- Published: 2025-04-28
- Snippet: SpaceX conducted its 49th Falcon 9 mission of 2025 and has another Starlink launch scheduled for later that night from Cape Canaveral Space Force Station.4. Amazon’s Satellite Internet Competition:
- Title: Amazon Just Launched Its First Internet Satellites to Compete Against SpaceX’s Starlink
- Source: The Associated Press
- Published: 2025-04-29
- Snippet: Amazon launched its first set of internet satellites using an Atlas V rocket, aiming to compete with SpaceX’s Starlink.5. Recent Starlink Satellite Launch:
- Title: SpaceX launches batch of Starlink satellites into low-Earth orbit
- Source: UPI
- Published: 2025-04-28
- Snippet: SpaceX launched a Falcon 9 rocket carrying nearly two dozen Starlink satellites into low-Earth orbit from Florida’s east coast.
Inserting rows into `tools`: 1 rows [00:00, 363.87 rows/s]
Inserted 1 row with 0 errors.Query: What’s the weather in San Francisco?
Answer: Based on the tool output provided, here’s the current weather in San Francisco:- Temperature: 72°F
- Condition: Partly CloudyIf you need more detailed information or a forecast for the upcoming days, feel free to ask!
How It Works
This notebook demonstrates two key Bedrock integration patterns:
-
Basic Message Completion: Using the
bedrock.messages()
function to generate responses from Bedrock models.
-
Tool-based Agent: Using
bedrock.converse() and
bedrock.invoke_tools() to create an agent that can:
- Analyze user queries
- Select appropriate tools (news search or weather)
- Execute the tools to retrieve information
- Generate a final response based on the tool output
This pattern demonstrates how Pixeltable can be used as an agent
engineering framework, allowing you to create complex, multi-step
workflows with LLMs and external tools.
Learn More
To learn more about advanced techniques like RAG operations in
Pixeltable, check out the RAG Operations in
Pixeltable
tutorial.
For more on agent engineering with Pixeltable, see the PixelAgent
documentation.
If you have any questions, don’t hesitate to reach out.
