opendcai/dataflow-webui/format-str-prompted-generator

FormatStrPromptedGenerator Operator Reference

FormatStrPromptedGenerator is DataFlow's multi-field template-based LLM generation operator. It maps multiple dataframe columns into template placeholders, builds one prompt per row, calls the LLM, writes the result into output_key, persists the dataframe, and returns the output_key string.

1. Imports

from dataflow.operators.core_text import FormatStrPromptedGenerator
from dataflow.prompts.core_text import FormatStrPrompt

2. Constructor

FormatStrPromptedGenerator(
    llm_serving,
    system_prompt="You are a helpful agent.",
    prompt_template=FormatStrPrompt(f_str_template="..."),
    json_schema=None,
)

Important prompt_template Notes

1. The source code default is FormatStrPrompt, which is the class object, not

an instance. Because of @prompt_restrict, omitting prompt_template entirely can raise TypeError at construction time.

1. Passing prompt_template=None is also invalid here. Unlike

BenchAnswerGenerator, this operator explicitly raises:

ValueError("prompt_template cannot be None")

1. The practical safe pattern is to always pass an instantiated template:

FormatStrPrompt(
    f_str_template="Title: {title}\n\nBody: {body}\n\nSummarize this article."
)

Constructing FormatStrPrompt

FormatStrPrompt(
    f_str_template="{title}\n\n{body}",
    on_missing="raise",
)

3. run() Signature

op.run(
    storage=self.storage.step(),
    output_key="generated_content",
    title="title_col",
    body="body_col",
)
# returns: output_key

Placeholder Mapping Rule

In run(...), each kwarg uses this convention:

• kwarg name = placeholder name inside f_str_template
• kwarg value = dataframe column name to read from

Example:

prompt_template = FormatStrPrompt(
    f_str_template="Title: {title}\n\nBody: {body}\n\nSummarize this article."
)
generator.run(
    storage=self.storage.step(),
    output_key="summary",
    title="headline_col",
    body="article_body_col",
)

This means:

• {title} is replaced with row["headline_col"]
• {body} is replaced with row["article_body_col"]

4. Actual Execution Logic

The current implementation behaves as follows:

1. Read the dataframe from storage.
2. Collect need_fields = set(input_keys.keys()).
3. For each row, build:

key_dict = {key: row[input_keys[key]] for key in need_fields}

1. Call:

prompt_text = prompt_template.build_prompt(need_fields, **key_dict)

1. Append all row prompts into llm_inputs.
2. Call:

llm_serving.generate_from_input(
    user_inputs=llm_inputs,
    system_prompt=self.system_prompt,
    json_schema=self.json_schema,
)

1. Write generated outputs into dataframe[output_key].
2. Persist via storage.write(dataframe).
3. Return output_key.

5. Important Rules

1. Always pass an instantiated FormatStrPrompt(...) or a compatible DIY prompt instance.
2. Do not omit prompt_template, do not pass the class object, and do not pass None.
3. Every value in **input_keys must be an existing dataframe column name.
4. The operator does not validate template placeholders against the template string itself. If you forget to map a placeholder that appears in f_str_template, that placeholder can remain unreplaced in the final prompt.
5. If you provide a placeholder name in **input_keys that is not used in the template, the extra value is simply unused by string replacement.
6. The operator adds or overwrites output_key, and does not filter rows.
7. The return value is the bare output_key string, not a list.

6. Typical Usage

from dataflow.operators.core_text import FormatStrPromptedGenerator
from dataflow.prompts.core_text import FormatStrPrompt
prompt_template = FormatStrPrompt(
    f_str_template="Title: {title}\n\nBody: {body}\n\nPlease write a concise summary."
)
generator = FormatStrPromptedGenerator(
    llm_serving=self.llm_serving,
    system_prompt="You are a professional editor. Write concise, informative summaries.",
    prompt_template=prompt_template,
)
generator.run(
    storage=self.storage.step(),
    output_key="summary",
    title="title",
    body="body",
)