Using AI in Workflows

Use the ai module for LLM completions in your workflows. It uses the platform default profile unless you pass profile= to select another reusable model profile.

Basic Completion

from bifrost import workflow, ai

@workflow
async def summarize(text: str):
    response = await ai.complete(
        f"Summarize: {text}",
        profile="Summarization",
    )
    return {"summary": response.content}

The response includes:

• content - Generated text
• input_tokens - Tokens in prompt
• output_tokens - Tokens in response
• model - Model used

If you omit profile=, Bifrost uses the profile assigned as Default under Settings → AI → Models. Use model= only when you intentionally want to override the model id stored on that profile.

Message Format

For multi-turn conversations or system prompts:

response = await ai.complete(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Explain Kubernetes."},
    ],
)

Or use the system parameter:

response = await ai.complete(
    "Explain Kubernetes.",
    system="You are a helpful assistant.",
)

Streaming Responses

For long responses, stream tokens as they’re generated:

from bifrost import ai

@workflow
async def stream_response(prompt: str):
    chunks = []
    async for chunk in ai.stream(prompt):
        chunks.append(chunk.content)
        if chunk.done:
            print(f"Total tokens: {chunk.input_tokens + chunk.output_tokens}")

    return {"response": "".join(chunks)}

Structured Outputs

Get responses as Pydantic models:

from pydantic import BaseModel
from bifrost import workflow, ai

class TicketAnalysis(BaseModel):
    priority: str
    category: str
    summary: str

@workflow
async def analyze_ticket(description: str):
    analysis = await ai.complete(
        f"Analyze this ticket: {description}",
        response_format=TicketAnalysis,
    )
    return analysis.model_dump()

Tip

Structured outputs use JSON schema generation to ensure the LLM returns valid data matching your model.

Model Overrides

Override the default model for specific calls:

# Use a specific model within the selected profile/provider
response = await ai.complete(
    "Complex task requiring reasoning",
    profile="Summarization",
    model="gpt-4o",
)

# Or a different reusable profile
response = await ai.complete(
    "Creative writing task",
    profile="Creative",
    model="claude-3-5-sonnet-latest",
)

With Knowledge Base (RAG)

Include context from your knowledge store:

response = await ai.complete(
    "What is our refund policy?",
    knowledge=["policies", "faq"],  # Namespace names
)

Bifrost automatically:

1. Searches the specified namespaces
2. Includes relevant documents as context
3. Generates a response grounded in your data

See Knowledge Bases for setup.

Parameters

Parameter	Type	Description
prompt	str	Simple prompt (alternative to messages)
messages	list	Chat messages with role/content
system	str	System prompt
response_format	type	Pydantic model for structured output
knowledge	list[str]	Knowledge namespaces for RAG
profile	str	Reusable model profile name for ai.complete(); omit for the default profile
model	str	Override default model
org_id	str	Organization context (auto-set in workflows)

Error Handling

from bifrost import ai

@workflow
async def safe_complete(prompt: str):
    try:
        response = await ai.complete(prompt)
        return {"result": response.content}
    except Exception as e:
        return {"error": str(e)}

Next Steps

• Knowledge Bases - RAG with hybrid semantic and full-text search
• Agents and Chat - Conversational AI