dspy-two-step-adapter

Use when working with reasoning models (o1, o3, o3-mini, DeepSeek-R1, Claude extended thinking) that reject system prompts or ignore formatting instructions. Common scenarios - using o1 or o3 with DSP

SKILL.md

GitHub

Use Reasoning Models with dspy.TwoStepAdapter

Guide the user through configuring DSPy to work with reasoning models (o1, o3, o3-mini, DeepSeek-R1, Claude with extended thinking) that need special handling for structured output.

Why reasoning models need TwoStepAdapter

Reasoning models (o1, o3, DeepSeek-R1, Claude extended thinking) behave differently from standard chat models:

  • They reject or ignore system prompts (o1/o3 strip them)
  • They ignore formatting instructions (the model "thinks" and produces free-form output)
  • They cannot follow ChatAdapter's field delimiters ([[ ## field_name ## ]])

TwoStepAdapter solves this with a two-phase approach:

  1. Phase 1 (main LM): The reasoning model generates freely -- no formatting constraints
  2. Phase 2 (extraction LM): A fast, cheap model parses the reasoning output into structured fields

When to use TwoStepAdapter

Use TwoStepAdapter when... Use ChatAdapter (default) when...
Using o1, o3, o3-mini Using GPT-4o, Claude, Gemini
Using DeepSeek-R1 Using any instruction-following model
Using Claude with extended thinking The model follows formatting reliably
Model ignores your output format Structured output works out of the box
Getting raw reasoning dumps instead of fields You do not need reasoning-heavy processing

Step 1: Basic TwoStepAdapter setup

import dspy

# The reasoning model (generates freely)
main_lm = dspy.LM("openai/o3-mini")

# The extraction model (parses into structured fields)
extraction_lm = dspy.LM("openai/gpt-4o-mini")

# Configure the adapter (only the extraction model is passed in)
adapter = dspy.TwoStepAdapter(extraction_model=extraction_lm)

# The reasoning model is the configured lm, not an adapter argument
dspy.configure(lm=main_lm, adapter=adapter)

Now use DSPy normally -- the adapter handles the two-phase flow transparently:

qa = dspy.ChainOfThought("question -> answer")
result = qa(question="What is 127 * 389?")
print(result.answer)  # Structured output, extracted by gpt-4o-mini

Step 2: Model pairing recommendations

Reasoning model (dspy.configure(lm=...)) Extraction model (extraction_model=...) Notes
openai/o1 openai/gpt-4o-mini Best reasoning + cheap extraction
openai/o3 openai/gpt-4o-mini Highest capability
openai/o3-mini openai/gpt-4o-mini Cost-effective reasoning
deepseek/deepseek-r1 openai/gpt-4o-mini Open-weight reasoning
anthropic/claude-sonnet-4-5-20250929 (extended thinking) anthropic/claude-haiku-3-5-20241022 All-Anthropic stack

Pairing principles:

  • The extraction LM should be fast and cheap -- it just parses structured fields from text
  • The extraction LM must follow formatting instructions well (ChatAdapter-compatible)
  • Cross-provider pairing is fine (o3 + Claude Haiku works)

Step 3: Per-module adapter assignment

You can use TwoStepAdapter for specific modules while using ChatAdapter elsewhere:

import dspy

# Default: fast model with ChatAdapter
fast_lm = dspy.LM("openai/gpt-4o-mini")
dspy.configure(lm=fast_lm)

# Reasoning model for hard problems
reasoning_lm = dspy.LM("openai/o3-mini")
extraction_lm = dspy.LM("openai/gpt-4o-mini")
reasoning_adapter = dspy.TwoStepAdapter(extraction_model=extraction_lm)


class Pipeline(dspy.Module):
    def __init__(self):
        # Simple classification -- uses default ChatAdapter
        self.classify = dspy.Predict("text -> category")

        # Hard reasoning -- uses TwoStepAdapter
        self.analyze = dspy.ChainOfThought("text, category -> analysis")
        self.analyze.adapter = reasoning_adapter
        self.analyze.lm = reasoning_lm

    def forward(self, text):
        category = self.classify(text=text).category
        analysis = self.analyze(text=text, category=category).analysis
        return dspy.Prediction(category=category, analysis=analysis)

Step 4: Handling extended thinking (Claude)

Claude with extended thinking uses a budget parameter instead of a separate model:

import dspy

# Claude with extended thinking enabled
thinking_lm = dspy.LM(
    "anthropic/claude-sonnet-4-5-20250929",
    thinking={"type": "enabled", "budget_tokens": 10000},
)

# Extraction model
extraction_lm = dspy.LM("anthropic/claude-haiku-3-5-20241022")

adapter = dspy.TwoStepAdapter(extraction_model=extraction_lm)

# The thinking model is the configured lm, not an adapter argument
dspy.configure(lm=thinking_lm, adapter=adapter)

# Works transparently -- thinking model reasons, haiku extracts structure
solver = dspy.ChainOfThought("problem -> solution")
result = solver(problem="Prove that sqrt(2) is irrational")

Step 5: When NOT to use TwoStepAdapter

Do not use TwoStepAdapter when:

  • The model already follows formatting -- standard models (GPT-4o, Claude, Gemini) work fine with ChatAdapter
  • You need minimum latency -- two LM calls instead of one adds overhead
  • The task is simple -- reasoning models are overkill for classification or extraction
  • Cost is critical -- you pay for both the reasoning call and the extraction call

If the model follows ChatAdapter format reliably, TwoStepAdapter adds cost and latency for no benefit.

Gotchas

  1. Claude uses ChatAdapter for o1/o3 models. Reasoning models reject ChatAdapter formatting. If you see raw reasoning dumps without structured fields, switch to TwoStepAdapter.
  2. Claude sets extraction_lm to the same reasoning model. The extraction model should be fast and cheap (gpt-4o-mini, Claude Haiku). Using o3 for extraction wastes money and is slower.
  3. Claude forgets to set both lm and adapter on per-module assignment. When assigning TwoStepAdapter to a specific module, set both module.lm = reasoning_lm and module.adapter = dspy.TwoStepAdapter(extraction_model=...). There is no main_lm constructor arg -- the reasoning model is the module's lm, and only the extraction model goes to the adapter. Missing either causes the wrong model or wrong adapter to be used.
  4. TwoStepAdapter is not needed for DeepSeek-V3. Only DeepSeek-R1 (the reasoning variant) needs it. DeepSeek-V3 follows formatting like a standard chat model.
  5. Claude wraps extended thinking in TwoStepAdapter incorrectly. For Claude extended thinking, pass thinking={"type": "enabled", "budget_tokens": N} to the LM constructor, not as a separate parameter.

Additional resources

Cross-references

Install any skill: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill <name>

  • All adapters overview -- see /dspy-adapters
  • ChatAdapter deep dive -- see /dspy-chatadapter
  • LM configuration and provider setup -- see /dspy-lm
  • Switching models without breaking prompts -- see /ai-switching-models
  • Install /ai-do if you do not have it -- it routes any AI problem to the right skill and is the fastest way to work: npx skills add lebsral/DSPy-Programming-not-prompting-LMs-skills --skill ai-do