Programmatic Pipelines (DSPy)

Expert

Build dynamic, code-driven prompt structures rather than static text files.

01 The Concept

Modern applications use programmatic pipelines (like DSPy or LangChain) where prompts are compiled dynamically from templates, few-shot examples are fetched based on embedding matches, and model routing is managed in code.

Instead of treating prompts as static text strings, programmatic pipelines treat them as components of an execution pipeline, dynamically compiling prompts based on task inputs.

02 Weak vs. Strong

EX 01Programmatic DSPy Pipeline

import dspy

Define structured signature

class QA(dspy.Signature): context = dspy.InputField(desc="retrieved documents") question = dspy.InputField() answer = dspy.OutputField(desc="fact-grounded summary")

Compile pipeline

qa_pipeline = dspy.ChainOfThought(QA) prediction = qa_pipeline(context=doc_list, question="what is X?")

→ Why it works

Treats prompting as a typed pipeline signature. The framework compiles variables and can automatically search for the best instructions based on training logs.

03 Key Points

01Prompts as Functions: Declare inputs, logic, and output schemas inside code structures.

02Dynamic Compilation: Construct prompts on-the-fly by stitching context, instructions, and history.

03Semantic Retrieval: Fetch matching few-shot examples from vector databases at runtime.

04DSPy Optimization: Let the framework auto-tune prompt templates and few-shot selections using code compilers.

05State Integration: Coordinate multi-step API routing, database updates, and tool calls.

04 Model-Specific Notes

Anthropic pipelines benefit from strict JSON schemas generated programmatically by pipelines to handle tool routing.

05 For Your Role

It's like building an automated factory line where materials (data) are assembled, checked, and boxed up by robots (code), rather than packing each box by hand.

Next: Prompt Ensembling via MoE →