mirror of
https://github.com/magnus919/agent-skills.git
synced 2026-09-17 22:46:29 +03:00
Magnus Hedemark
6c193ae2d8
DSPy v1.1.0: validation audit, worked RAG compilation example, expand ref table Haystack v1.1.0: validation audit, file converters/YAML/component types, +2 refs CrewAI v1.1.0: validation audit, unified Memory system, Flows docs, +3 refs AutoGen v1.1.0: validation audit, v0.4 migration guide, AgentTool, streaming, +2 refs All API surfaces validated against official docs.
77 lines
2.3 KiB
Markdown
77 lines
2.3 KiB
Markdown
# DSPy — Worked Example: Full RAG Compilation
|
|
|
|
This example shows a complete DSPy program from definition through compilation, with expected output annotations.
|
|
|
|
## Program Definition
|
|
|
|
```python
|
|
import dspy
|
|
from dspy.datasets import DataLoader
|
|
|
|
lm = dspy.LM("openai/gpt-4o-mini")
|
|
dspy.configure(lm=lm)
|
|
|
|
class RAG(dspy.Module):
|
|
def __init__(self, k=3):
|
|
self.retrieve = dspy.Retrieve(k=k)
|
|
self.generate = dspy.ChainOfThought("context, question -> answer")
|
|
|
|
def forward(self, question):
|
|
context = "\n".join(self.retrieve(question).passages)
|
|
return self.generate(question=question, context=context)
|
|
```
|
|
|
|
## Dataset
|
|
|
|
```python
|
|
trainset = [
|
|
dspy.Example(question="What is DSPy?", answer="A compiler for prompt programs.").with_inputs("question"),
|
|
dspy.Example(question="What is a signature?", answer="Input/output field pairs defining a task.").with_inputs("question"),
|
|
dspy.Example(question="What is MIPROv2?", answer="Bayesian optimizer for joint instruction and demo tuning.").with_inputs("question"),
|
|
]
|
|
```
|
|
|
|
## Compilation
|
|
|
|
```python
|
|
from dspy.teleprompt import MIPROv2
|
|
|
|
def correct(example, pred, trace=None):
|
|
return example.answer in pred.answer
|
|
|
|
optimizer = MIPROv2(metric=correct, auto="light")
|
|
compiled = optimizer.compile(RAG(), trainset=trainset)
|
|
```
|
|
|
|
**Expected compile output:**
|
|
- Compiler output logs showing: bootstrapping demos, proposing instruction candidates, evaluating candidates, selecting best
|
|
- Typical run: ~30-60 seconds, ~150-300 API calls (auto="light")
|
|
- Output: a compiled module with `_compiled = True` flag set
|
|
|
|
## Inference
|
|
|
|
```python
|
|
# Use the compiled program
|
|
result = compiled(question="What is DSPy compiling?")
|
|
|
|
# Expected result structure:
|
|
print(result) # dspy.Prediction object
|
|
print(result.answer) # The generated answer string
|
|
# ChainOfThought also provides:
|
|
print(result.reasoning) # The reasoning chain used
|
|
|
|
# Save for deployment
|
|
compiled.save("rag_program.json")
|
|
|
|
# Later, reload
|
|
loaded_rag = RAG()
|
|
loaded_rag.load("rag_program.json")
|
|
```
|
|
|
|
## Expected Output Depth
|
|
|
|
- Uncompiled: Answer based solely on LM training data, no optimization
|
|
- BootstrapFewShot: Higher quality, uses successful traces as demos
|
|
- MIPROv2: Highest quality, optimized instructions + demos together
|
|
- Cost: ~$0.50-2.00 for auto="light" on gpt-4o-mini
|