LLM Chain Processing Example¶

Note: LLM functionality has moved to the dedicated dataknobs-llm package.

This example has been superseded by comprehensive conversation flow examples in the LLM package.

New Location¶

For LLM chain processing and multi-step workflows, see:

Conversation Flow Examples - Multi-step conversation patterns
Advanced Prompting Examples - Prompt chaining and RAG
FSM-Based Conversation Flow - FSM orchestration

Supported Patterns¶

The LLM package now provides these patterns directly:

Sequential LLM Chains¶

Use conversation flows for sequential processing:

from dataknobs_llm.conversations.flow import ConversationFlow, FlowState, KeywordCondition

chain_flow = ConversationFlow(
    name="sequential_chain",
    initial_state="summarize",
    states={
        "summarize": FlowState(
            prompt="summarize_text",
            transitions={"next": KeywordCondition([".*"])},
            next_states={"next": "analyze"}
        ),
        "analyze": FlowState(
            prompt="analyze_summary",
            transitions={"next": KeywordCondition([".*"])},
            next_states={"next": "conclude"}
        ),
        "conclude": FlowState(
            prompt="draw_conclusions",
            transitions={},
            next_states={}
        )
    }
)

Chain-of-Thought Reasoning¶

Implement step-by-step reasoning flows:

cot_flow = ConversationFlow(
    name="chain_of_thought",
    initial_state="decompose",
    states={
        "decompose": FlowState(
            prompt="break_down_problem",
            transitions={"decomposed": KeywordCondition([".*"])},
            next_states={"decomposed": "solve_steps"}
        ),
        "solve_steps": FlowState(
            prompt="solve_each_step",
            transitions={"solved": KeywordCondition([".*"])},
            next_states={"solved": "synthesize"}
        ),
        "synthesize": FlowState(
            prompt="combine_solutions",
            transitions={},
            next_states={}
        )
    }
)

RAG (Retrieval-Augmented Generation)¶

Configure RAG in prompt templates:

# prompts/user/rag_query.yaml
template: |
  Answer this question using the provided context:

  Question: {{question}}

  Context:
  {{RAG_DOCS}}

rag_configs:
  - adapter_name: knowledge_base
    query_template: "{{question}}"
    k: 5
    placeholder: "RAG_DOCS"

Prompt Chaining¶

Chain prompts with context preservation:

# First prompt
await manager.add_message(
    role="user",
    prompt_name="initial_analysis",
    params={"data": raw_data}
)
result1 = await manager.complete()

# Second prompt - context automatically included
await manager.add_message(
    role="user",
    prompt_name="deep_dive",
    params={"aspect": "key_findings"}
)
result2 = await manager.complete()  # Has context from result1