← Back to Dashboard
1. InvokeModel and Streaming Basics2. Converse API Patterns3. Five Ways to Call Bedrock: Converse, Invoke, Messages, Responses, Chat Completions
Converse API Patterns
📚 Inference APIs⏱ 11 min⭐ 70 XP
Message-First Orchestration
The Converse API unifies multi-turn interactions and tool-ready structures for many Bedrock-supported model families. One request shape works across Amazon Nova, Anthropic Claude, Meta Llama, Mistral, and more - swap the modelId and your code keeps working.
Anatomy of a Converse Request
response = client.converse(
modelId="amazon.nova-pro-v1:0",
system=[{"text": "You are a support triage assistant. Reply in strict JSON."}],
messages=[
{"role": "user", "content": [{"text": "Customer reports checkout 500 errors since 09:40 UTC"}]},
{"role": "assistant", "content": [{"text": '{"severity":"P1","team":"payments"}'}]},
{"role": "user", "content": [{"text": "They add: only EU users affected"}]}
],
inferenceConfig={"maxTokens": 400, "temperature": 0.1, "stopSequences": ["###"]},
)
| Field | Purpose | Production guidance |
|---|---|---|
system | Policy, persona, output rules - kept out of user-editable content | Version it; keep stable across turns |
messages | Alternating user/assistant turns; content is a list of blocks (text, image, document) | You own history - the API is stateless, so trim/summarise long conversations |
inferenceConfig | maxTokens, temperature, topP, stopSequences | Low temperature (0–0.3) for deterministic business tasks |
toolConfig | Tool/function definitions for agentic loops | Covered in the Tool Use module |
Design Each Turn With
- Stable system instructions for policy and tone - changing them mid-conversation causes behavioral drift.
- User content with clear constraints - format, length, and refusal expectations.
- Tool invocation hooks (when applicable) via
toolConfig. - Structured post-processing and safety checks - validate JSON against a schema before acting on it.
Key insight: Converse is stateless - Bedrock does not remember previous turns. Your application sends the full relevant history every call, which means you control context cost: prune, summarise, or window old turns deliberately.
🧪 Knowledge Check
Press 1-4 to select1 of 4
What is a key benefit of the Converse-style API?
No need for IAM
Consistent message-based workflow for multi-turn interactions
Only supports one model
Removes all safety requirements