AI Tools
Dspy
Mental Model

DSPy Mental Model

Glossary

  • [LM prompt] Pipeline
    • a stack of LM prompting tecnhiques, that solves a complex task
    • a stack of LM prompts
  • Prompt template
    • a string
  • Text transformation graph
    • Imperative computation graph
  • Declarative module
  • Parameterized module
  • Composition
  • Prompting
  • Fine-tuning
  • Augmentation
  • Reasoning
  • Compiler
  • Metric
  • Multi-hop retrieval
  • Agent loops
  • Self-bootstrap [pipelines]
  • Few-shot prompting
  • Demonstrations
    • [expert-created]
  • Prompt-chains
    • [expert-written]

As can be seen based on the paper title, it is about:

  • Entities
    • LM calls
    • Pipelines
    • Compilation

So, it is about compiling LM calls into pipelines.

To solve complex tasks there are:

  • LM prompting techniques

    • Stacks of those (Pipelines)

  • Complex tasks are solved by stacking LM prompts into pipelines.

  • Existing [LM prompt] pipelines typically use hard-coded prompt templates (strings)

  • DSPy is to develop and optimize [LM prompt] pipelines

    • by providing a more systematic approach to do it

  • DSPy abstracts [LM prompt] pipelines as text transformation graphs

    • LMs are invoked through declarative modules

  • DSPy modules are parameterized

  • DSPy modules can learn

    • (by creating and collecting demonstrations)
    • how to apply compositions of
      • prompting [techniques]
      • finetuning [techniques]
      • augmentation [techniques]
      • reasoning [techniques]
DspySdk