Build LLM apps as explicit flowcharts, not black boxes

What it does

LLMFlows is a Python framework for chaining Large Language Model calls into structured workflows. It provides thin wrappers around APIs like OpenAI, plus PromptTemplate, MessageHistory, and FlowStep classes that pass outputs between dependent steps. The library runs steps only when their inputs are ready and can execute independent branches in parallel via async flows.

The interesting bit

The whole pitch is transparency: every flowstep exposes its exact prompt, model configuration, raw API response, token count, retries, and timing. That turns the usual black-box LLM orchestration into something you can actually debug in production without guessing what hidden template generated a query.

Key highlights

• Explicit flows: Connect FlowStep nodes into directed graphs; the Flow class resolves dependencies and passes outputs as inputs to downstream steps.
• Async execution: AsyncFlowStep runs independent branches in parallel when their inputs are ready, cutting runtime for complex pipelines.
• Built-in observability: Each step returns call metadata, raw API responses, token usage, retry counts, and timing—no hidden prompts or magic strings.
• Vector store support: VectorStoreFlowStep integrates with databases like Pinecone for retrieval-augmented generation inside flows.
• Callback hooks: Inject functions at different stages for custom logging, tracing, or side effects.

Verdict

Best for developers who need to trace, retry, and wire together OpenAI-centric workflows with strict observability. Less compelling if you are looking for a generic, provider-agnostic orchestrator with broad third-party integrations.