Talks - Tuana Celik: Everything is a graph, including LLM Applications (and that’s handy)

AI/LLM applications can be effectively modeled and implemented as interconnected graphs composed of specialized components

Components in the graph are responsible for individual tasks (fetching, embedding, generating, classifying, etc.) and pass data from one node to another

Haystack 2.0 provides a framework for building these graph-based pipelines while allowing custom components based on specific needs

Pipeline graphs can be:

• Linear (simple sequence of tasks)
• Branching (parallel processing paths)
• Cyclical (loops for refinement/validation)

Large Language Models don’t need to handle every task - specialized components can be more efficient for specific operations like translation or classification

Structured output validation (using tools like Pydantic) helps ensure LLM responses match expected formats

Components should have clear input/output types that match the requirements of connected nodes in the pipeline

Complex AI applications can be broken down into smaller, manageable tasks connected through a pipeline graph

Pipeline approach allows easy swapping of components (e.g., switching between different LLM providers)

Error handling and validation can be built into the pipeline to allow for automatic retries and refinement of results