Generating text with language models

Introduction to Amazon Bedrock

Nikhil Rangarajan

Data Scientist

Sending prompts to language models

Prompt: the input text you send to the model
- "Explain how AWS Lambda works"

An icon representing the prompt.

Sending prompts to language models

Prompt: the input text you send to the model
- "Explain how AWS Lambda works"
Completion: the text generated by the model
- "AWS Lambda is a serverless compute service that..."

Icons representing the prompt and completion.

Sending prompts to language models

Prompt: the input text you send to the model
- "Explain how AWS Lambda works"
Completion: the text generated by the model
- "AWS Lambda is a serverless compute service that..."
Response:
- Generated text (completion)
- Metadata (tokens, model info)
- Status information

Three icons in sequence representing the prompt, completion, and response.

Applying basic prompt engineering techniques

Techniques for better results:
- Clarity and specificity: write detailed and precise instructions
- Role assignment: give a persona to improve contextual understanding
- Formatting instructions: add constraints to guide the response

"Explain the benefits of exercise 
in simple terms."

"You are a nutritionist. Explain the 
benefits of a balanced diet."

"List the top 3 benefits of cloud 
computing in bullet points."

Advanced text parsing techniques

Handle missing or unexpected keys

data = json.loads(response['body'].read())
if 'completion' in data:
    output = data['completion']

else:
    output = "Key not found"

Multi-Level JSON Parsing to navigate deeply nested structures

try:
    nova_output = data.decode()["output"]

except (KeyError, IndexError) as e:
    nova_output = f"Error: {str(e)}"

Response processing and text handling

Unicode or encoding issues: handle special characters in text responses

# Ensure proper encoding
output = data['completion'].encode('utf-8').decode('unicode_escape')
print(output)

Processing large responses: split or truncate large outputs for usability

# Truncate response for display
output = data['completion'][:500]  # Limit to 500 characters
print(output)

Category extraction with Bedrock

Example: classifying text data
Use structured prompts with clear category options
Provide explicit instructions for single-category output
Process model responses efficiently

Result: Accurate text categorization

Icons representing text classification

Category extraction with Bedrock

def categorize_ticket(ticket, categories):
  prompt = f"Categorize into one: {', '.join(categories)}. 
  Ticket: {ticket}"
  response = bedrock.invoke_model(
    modelId='amazon.nova-micro-v1:0',
    body=json.dumps({"messages": [{"role": "user", "content": 
                                   [{"text": prompt}]}]}))

  return json.loads(nova_response.get("body").read().decode())["output"]

{"output": {"message": {"role": "assistant","content": [
        {
          "text": "Customer Service"
        }]}, "stop_reason": "stop_sequence"}}

Let's practice!

Introduction to Amazon Bedrock