dspy-template-adapter 0.2.0

A DSPy Adapter for exact-fidelity prompt templates with full control over messages.

dspy-template-adapter

A DSPy Adapter that gives you exact control over the messages sent to the LM — no hidden prompt rewriting. Define your prompt as a list of messages (just like the OpenAI API), and the adapter handles variable interpolation, few-shot demos, conversation history, output parsing, and integration with DSPy's full optimization pipeline.

pip install dspy-template-adapter

Why?

DSPy's built-in adapters (ChatAdapter, JSONAdapter) rewrite your prompt into DSPy's own scaffolding format ([[ ## field ## ]] markers, field descriptions, etc.). This is great for optimization, but it means you can't reproduce the exact API calls from a vanilla OpenAI/Anthropic prompt.

TemplateAdapter solves this: your messages are the prompt. The signature defines the I/O contract, the adapter renders your templates, and DSPy handles the rest (caching, tracing, retries, evaluation, optimization).

Quickstart

import dspy
from dspy_template_adapter import TemplateAdapter, Predict

# 1. Define a signature (the I/O contract)
class Summarize(dspy.Signature):
    """Summarize input text concisely."""
    text: str = dspy.InputField()
    summary: str = dspy.OutputField()

# 2. Define the prompt template
adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "You are a concise assistant. {instruction}"},
        {"role": "user", "content": "Summarize:\n\n{text}"},
    ],
    parse_mode="full_text",
)

# 3. Bind adapter to predictor
summarizer = Predict(Summarize, adapter=adapter)

# 4. Configure LM and call
dspy.configure(lm=dspy.LM("gpt-4.1-nano"))
out = summarizer(text="DSPy is a framework for programming language models.")
print(out.summary)

What happens under the hood:

1. {instruction} is replaced with the signature docstring ("Summarize input text concisely.")
2. {text} is replaced with the input value
3. The two rendered messages are sent to the LM — nothing else added
4. The full response is mapped to summary (because parse_mode="full_text")

Core Concepts

The adapter has two jobs:

• format() — Render your message templates + input values into the final messages list sent to the LM
• parse() — Extract output fields from the LM's raw completion string

Template Syntax

Inside message content strings, you can use:

Syntax	What it does
{field_name}	Replaced with the value of that input field
{instruction}	Replaced with the signature's docstring
{inputs()}	Renders all input fields. Supports style='yaml', style='json', style='xml', or default
{outputs()}	Renders output field descriptions. Supports style='schema', style='xml' (+ optional wrap), or default
{demos()}	Renders few-shot demos inline. Supports style='json', style='yaml', style='xml', or default
{my_helper()}	Calls a custom function registered with register_helper()
{{ / }}	Literal braces (escaped)

Two special directive roles expand into multiple messages at render time:

Directive	What it does
{"role": "demos"}	Expands into user/assistant message pairs for each demo
{"role": "history"}	Expands a dspy.History object into prior conversation turns

Parse Modes

Value	When to use	Requirement
"full_text"	Entire LM response is your output	Exactly 1 output field
"json" (default)	LM returns JSON with keys matching output fields	Any number of output fields
"xml"	LM returns <field_name>value</field_name> tags	Tags for every output field
"chat"	Delegates to DSPy's ChatAdapter parser	DSPy's [[ ## field ## ]] format
A callable	Custom extraction logic	(signature, completion) -> dict

Structured JSON Output

When your signature has multiple output fields:

class Triage(dspy.Signature):
    ticket: str = dspy.InputField()
    category: str = dspy.OutputField()
    priority: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Return JSON only with keys: category, priority."},
        {"role": "user", "content": "{inputs(style='yaml')}"},
    ],
    parse_mode="json",
)

triage = Predict(Triage, adapter=adapter)
out = triage(ticket="Production checkout failing for VIP users")
print(out.category, out.priority)

The JSON parser uses json_repair for robustness — it handles malformed JSON and can extract JSON objects embedded in surrounding text.

XML Output

Some models (especially Claude) perform well with XML-tagged output.

Hardcoded XML (explicit control)

class Review(dspy.Signature):
    text: str = dspy.InputField()
    sentiment: str = dspy.OutputField()
    reasoning: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": (
            "Analyze the sentiment. Respond with XML tags:\n"
            "<reasoning>your reasoning</reasoning>\n"
            "<sentiment>positive, negative, or neutral</sentiment>"
        )},
        {"role": "user", "content": "{text}"},
    ],
    parse_mode="xml",
)

reviewer = Predict(Review, adapter=adapter)
out = reviewer(text="This product exceeded all my expectations!")
print(out.sentiment, out.reasoning)

Signature-driven XML (build the prompt from the signature)

Instead of repeating field names and descriptions in the template, use {outputs(style='xml')} and {inputs(style='xml')} to generate the XML structure from the signature's metadata:

class TranslateToFrench(dspy.Signature):
    """You are an English-to-French translator. You only translate."""

    user_english_text: str = dspy.InputField(desc="The English text to translate")
    user_french_target: str = dspy.InputField(desc="The target French variant")

    is_translation_task: str = dspy.OutputField(desc="yes or no")
    corrected_english: str = dspy.OutputField(desc="the English input with fixes")
    detected_tone: str = dspy.OutputField(desc="the tone of the English text")
    french_variant: str = dspy.OutputField(desc="the French variant used")
    translation: str = dspy.OutputField(desc="the final French translation")

adapter = TemplateAdapter(
    messages=[
        {
            "role": "system",
            "content": (
                "{instruction}\n\n"
                "Respond ONLY with the following XML structure (no other text):\n"
                "{outputs(style='xml', wrap='response')}\n\n"
                "Rules:\n"
                "- First decide if this is a translation task.\n"
                "- Correct the English first, then translate.\n"
                "- Respect the French variant: "
                "<user_french_target>{user_french_target}</user_french_target>."
            ),
        },
        {
            "role": "user",
            "content": "{inputs(style='xml')}",
        },
    ],
    parse_mode="xml",
)

The {outputs(style='xml', wrap='response')} generates the XML schema block from the output field descriptions, and {inputs(style='xml')} wraps each input value in its field-name tag. If you add, remove, or rename a field in the signature, the prompt updates automatically.

XML parsing handles tags anywhere in the response, multiline values, and raises clear errors for missing fields. XML avoids the common problem of models escaping quotes inside JSON string values.

When parse_mode="xml", auto-injected demo assistant messages also use XML format (<field>value</field> tags) so the LM sees consistent formatting throughout the conversation.

Template Functions

{inputs()} — render all input field values

# Default (same as yaml): "field: value" lines
{"role": "user", "content": "{inputs()}"}

# JSON object
{"role": "user", "content": "{inputs(style='json')}"}

# XML tags: <field_name>value</field_name>
{"role": "user", "content": "{inputs(style='xml')}"}

{outputs()} — render output field descriptions

# Numbered list of field names/types
{"role": "system", "content": "Produce:\n{outputs()}"}

# JSON schema
{"role": "system", "content": "Match this schema:\n{outputs(style='schema')}"}

# XML tags with field descriptions as placeholder values
{"role": "system", "content": "Respond with:\n{outputs(style='xml')}"}

# XML tags wrapped in a root element (indented)
{"role": "system", "content": "Respond with:\n{outputs(style='xml', wrap='response')}"}

The xml style pulls descriptions from dspy.OutputField(desc="…"). For example, a signature with:

translation: str = dspy.OutputField(desc="the final French translation")
detected_tone: str = dspy.OutputField(desc="the tone of the English text")

renders {outputs(style='xml', wrap='response')} as:

<response>
  <translation>the final French translation</translation>
  <detected_tone>the tone of the English text</detected_tone>
</response>

{demos()} — render few-shot examples inline

# Demos as JSON objects inside a message
{"role": "system", "content": "Examples:\n{demos(style='json')}"}

# Demos as XML tags
{"role": "system", "content": "Examples:\n{demos(style='xml')}"}

Styles: 'json', 'yaml', 'xml', or default (numbered text blocks).

Few-Shot Demos: Three Strategies

A) Inline — demos as text inside a message

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Examples:\n{demos(style='json')}"},
        {"role": "user", "content": "{inputs(style='yaml')}"},
    ],
    parse_mode="json",
)

Result: 2 messages. Demos are text inside the system message.

B) Directive — demos as separate conversation turns

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Classify tickets."},
        {"role": "demos"},
        {"role": "user", "content": "{inputs(style='yaml')}"},
    ],
    parse_mode="json",
)

Each demo becomes a user + assistant message pair. You can customize the format:

{"role": "demos", "user": "Ticket: {ticket}", "assistant": "{outputs_json}"}

C) Auto-injection

If your template doesn't mention demos at all, they're automatically injected as user/assistant pairs before the final user message when an optimizer adds them. This ensures compatibility with DSPy's optimization pipeline (BootstrapFewShot, etc.) without template changes.

The {instruction} Slot

class Summarize(dspy.Signature):
    """Summarize input text concisely."""  # <-- this is {instruction}
    text: str = dspy.InputField()
    summary: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "You are helpful. {instruction}"},
        {"role": "user", "content": "{text}"},
    ],
    parse_mode="full_text",
)

DSPy optimizers like MIPRO and COPRO rewrite the signature's instruction string. Including {instruction} in your template lets them optimize your prompt. Omitting it makes your prompt fully static.

Conversation History

For multi-turn chatbots, use dspy.History and the {"role": "history"} directive:

class ChatSig(dspy.Signature):
    question: str = dspy.InputField()
    history: dspy.History = dspy.InputField()
    answer: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "You are a helpful chatbot."},
        {"role": "history"},
        {"role": "user", "content": "{question}"},
    ],
    parse_mode="full_text",
)

chat = Predict(ChatSig, adapter=adapter)
history = dspy.History(messages=[
    {"question": "What is 1+1?", "answer": "2"},
])
resp = chat(question="What is 2+2?", history=history)

The directive expands each history entry into user/assistant message pairs. If omitted, history is auto-injected before the last user message.

Image Support

The adapter works seamlessly with dspy.Image inputs. Images are automatically converted to multimodal content blocks (the image_url format the OpenAI API expects).

Single Image

from dspy.adapters.types import Image

class Describe(dspy.Signature):
    """Describe the image."""
    image: Image = dspy.InputField()
    description: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "You describe images in one sentence."},
        {"role": "user", "content": "What is in this image? {image}"},
    ],
    parse_mode="full_text",
)

dspy.configure(lm=dspy.LM("gpt-4.1-nano"))
describer = Predict(Describe, adapter=adapter)
img = Image.from_file("photo.png")
out = describer(image=img)
print(out.description)

The {image} placeholder is first rendered to DSPy's internal image marker, then split_message_content_for_custom_types() splits the user message into proper content blocks:

# What the LM actually receives:
[
    {"role": "system", "content": "You describe images in one sentence."},
    {"role": "user", "content": [
        {"type": "text", "text": "What is in this image? "},
        {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}}
    ]}
]

Multiple Images

from dspy.adapters.types import Image

class Compare(dspy.Signature):
    """Compare two images."""
    image_a: Image = dspy.InputField()
    image_b: Image = dspy.InputField()
    comparison: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Compare images in one sentence."},
        {"role": "user", "content": "Image A: {image_a}\nImage B: {image_b}\nCompare them."},
    ],
    parse_mode="full_text",
)

comparer = Predict(Compare, adapter=adapter)
out = comparer(
    image_a=Image.from_file("red.png"),
    image_b=Image.from_file("blue.png"),
)
print(out.comparison)

Each {image_*} placeholder becomes its own image_url block, with text blocks for the surrounding content.

Images with Other Parse Modes

Images work with any parse mode — full_text, json, xml, or custom:

from dspy.adapters.types import Image

class Analyze(dspy.Signature):
    """Analyze an image."""
    image: Image = dspy.InputField()
    question: str = dspy.InputField()
    answer: str = dspy.OutputField()
    confidence: str = dspy.OutputField()

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Answer the question about the image. Return JSON with keys: answer, confidence."},
        {"role": "user", "content": "{question}\n{image}"},
    ],
    parse_mode="json",
)

dspy.configure(lm=dspy.LM("gpt-4.1-nano"))
analyzer = Predict(Analyze, adapter=adapter)
out = analyzer(image=Image.from_file("photo.png"), question="What color is this?")
print(out.answer, out.confidence)

Custom Template Helpers

Register custom template functions for complex rendering logic:

def format_as_xml(ctx, signature, demos, **kwargs):
    tag = kwargs.get("tag", "input")
    parts = []
    for name in signature.input_fields:
        val = ctx.get(name, "")
        parts.append(f"<{tag}_{name}>{val}</{tag}_{name}>")
    return "\n".join(parts)

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Process XML input."},
        {"role": "user", "content": "{format_as_xml(tag='field')}"},
    ],
    parse_mode="full_text",
)
adapter.register_helper("format_as_xml", format_as_xml)

Helper signature: (ctx: dict, signature, demos: list, **kwargs) -> str

Per-Module Adapter Isolation

The Predict wrapper lets each predictor use a different adapter:

formal = Predict(Summarize, adapter=TemplateAdapter(
    messages=[{"role": "system", "content": "Formal tone."}, {"role": "user", "content": "{text}"}],
    parse_mode="full_text",
))
casual = Predict(Summarize, adapter=TemplateAdapter(
    messages=[{"role": "system", "content": "Casual tone."}, {"role": "user", "content": "{text}"}],
    parse_mode="full_text",
))

# Each call uses its own system prompt
out_formal = formal(text="Quantum computing is...")
out_casual = casual(text="Quantum computing is...")

Custom Parse Functions

Pass any callable as parse_mode:

import re

def extract_rating(signature, completion):
    match = re.search(r"(\d+)/10", completion)
    return {"rating": match.group(1) if match else "0"}

adapter = TemplateAdapter(
    messages=[
        {"role": "system", "content": "Rate quality 1-10. Format: N/10."},
        {"role": "user", "content": "{text}"},
    ],
    parse_mode=extract_rating,
)

Debugging

Preview without calling the LM

msgs = adapter.preview(Summarize, inputs={"text": "hello"})
for msg in msgs:
    print(f"[{msg['role']}] {msg['content']}")

Inspect what was actually sent

last = dspy.settings.lm.history[-1]
for msg in last["messages"]:
    print(f"[{msg['role']}] {msg['content']}")

Checklist

1. Import error? Ensure dspy-template-adapter is installed (pip install dspy-template-adapter)
2. Template wrong? Use adapter.preview(...) — no LM call needed
3. Parse error on full_text? Signature must have exactly 1 output field
4. Parse error on json? LM response must contain all output field keys
5. Parse error on xml? LM response must contain <field>...</field> for every output field
6. Demos missing? {demos()} inline and {"role": "demos"} directive are mutually exclusive with auto-injection — pick one
7. Literal braces eaten? Use {{ and }}

Finetuning Data Export

ft = adapter.format_finetune_data(
    Summarize,
    demos=[],
    inputs={"text": "DSPy is a framework."},
    outputs={"summary": "A framework for LMs."},
)
# ft["messages"] is an OpenAI-compatible message list with assistant response appended

Live Compatibility Tests (Optimizers + Agents)

This repo includes live integration tests that verify TemplateAdapter works with DSPy optimizers and agentic modules.

Covered:

• BootstrapFewShot
• MIPROv2
• GEPA (when available in your DSPy version)
• ReAct
• CodeAct (when supported by your DSPy version/runtime)
• RML availability check (explicit skip if not present)

Run them with uv:

RUN_LIVE_DSPY=1 uv run python -m pytest -q dspy_template_adapter/tests/test_live_dspy_compat.py -s

Notes:

• These tests are skipped by default (they need network/API keys).
• They use gpt-4.1-nano for consistency.
• On DSPy versions where a feature is not present (for example GEPA or RML), tests skip with an explicit reason.

Deep optimizer/message audit (JSON output)

For a full before/after audit (accuracy, instruction changes, and raw message payloads from lm.history) run:

uv sync
uv run python scripts/live_optimizer_audit.py

Outputs are written under:

artifacts/optimizer_audit/YYYYMMDD-HHMMSS/

Detailed guide: docs/optimizer-audit.md

API Reference

TemplateAdapter(
    messages: list[dict],         # Message templates (required)
    parse_mode: str | callable,   # "json" (default), "full_text", "xml", "chat", or callable
)

Methods:
    .format(signature, demos, inputs) -> list[dict]
        Render messages from templates.

    .parse(signature, completion) -> dict
        Extract output fields from LM response.

    .preview(signature, demos=None, inputs=None) -> list[dict]
        Render without calling LM. For debugging.

    .register_helper(name, fn)
        Register custom {name()} template function.
        fn: (ctx, signature, demos, **kwargs) -> str

    .format_finetune_data(signature, demos, inputs, outputs) -> dict
        Generate OpenAI-compatible finetuning entry.

Predict(signature, adapter=adapter)
    dspy.Predict subclass with per-module adapter binding.

License

MIT