What are LLM Workflows?

While interacting with a Large Language Model (LLM) often starts with a simple prompt-and-response exchange, building useful applications requires more structure. Think past a single query. Real-world tasks often involve multiple steps, data transformations, and potentially several interactions with an LLM or other services.

An LLM workflow refers to this structured sequence of operations designed to accomplish a specific objective using one or more LLMs as components. It's the process of orchestrating inputs, LLM calls, data processing, and potentially external tools or data sources to produce a desired outcome.

Consider these points:

• More than Single Calls: While a basic interaction involves sending a prompt and receiving a response, a workflow might involve preprocessing user input, retrieving relevant information from a database, constructing a detailed prompt using that information, calling the LLM, parsing the structured output, and then taking further action based on that output.
• Orchestration: Workflows manage the flow of data and control between different components. This could be as simple as passing the output of one step as the input to the next, or it could involve complex conditional logic and decision-making, sometimes even using the LLM itself to decide the next step.
• Building Blocks for Applications: Workflows are the fundamental structures used to integrate LLM capabilities into functional software. Whether you're building a chatbot, a document summarizer, a code generator, or a data analysis tool, you are essentially designing and implementing an LLM workflow.

Simple vs. Complex Workflows

Workflows can range significantly in complexity:

• Simple Workflow: A straightforward example might involve taking user text, using a template to format it into a specific prompt (e.g., "Summarize the following text: {user_text}"), sending it to an LLM, and displaying the raw summary.
• Complex Workflow: A more involved workflow, like one for Retrieval-Augmented Generation (RAG), might look like this:
  1. Receive a user question.
  2. Use the question to search a vector database for relevant document chunks.
  3. Combine the question and the retrieved context into a prompt.
  4. Send the combined prompt to an LLM.
  5. Receive the LLM's response, which is grounded in the provided context.
  6. Format the response for the user.

This diagram illustrates a generic structure:

A typical flow involves processing input, preparing a prompt, interacting with the LLM, and handling the output. Some workflows may involve feedback loops or multiple LLM calls.

Understanding workflows is fundamental because it shifts the perspective from treating the LLM as a standalone magic box to seeing it as a powerful component within a larger, engineered system. This course focuses on using Python and specific libraries like LangChain and LlamaIndex to design, implement, test, and deploy these workflows effectively.