Skip to content

Data generation

QAPairGenerator

Bases: BaseModel

Generates questions and answers for sections and subsections of a document.

Parameters:

Name Type Description Default
openai_key str

The API key for OpenAI.

 required
openai_model str

The name of the OpenAI model to use.

 required
seed int

The seed for the random number generator. Defaults to 42.

 required
Source code in docqa/core/data_generation.py 
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
class QAPairGenerator(BaseModel):
    """
    Generates questions and answers for sections and subsections of a document.

    Args:
        openai_key (str): The API key for OpenAI.
        openai_model (str): The name of the OpenAI model to use.
        seed (int, optional): The seed for the random number generator. Defaults to 42.
    """

    class Config:
        arbitrary_types_allowed = True

    openai_key: str
    openai_model: str
    seed: int = 42

    @computed_field  # type: ignore[misc]
    @property
    def openai_client(self) -> OpenAI:
        return OpenAI(api_key=self.openai_key)

    output_format: str = (
        "Present your questions along with the detailed answers in the following JSON"
        ' format: [{"question": str, "answer": str}, ...].'
    )

    @computed_field  # type: ignore[misc]
    @property
    def question_types(self) -> dict[str, dict[str, str]]:
        return {
            "sparse": {
                "system_message": (
                    "You are a professional examiner. Your job is to give questions to"
                    " test people's understanding of a given document."
                ),
                "instruction": (
                    "You are given a document, your goal is:\n- construct a list of"
                    " different complex questions that can be answered based **solely**"
                    " on the given text.\n- make sure to cover all of the topics"
                    " described in the document.\n- include the answer for each"
                    " question, the answers should be as detailed as"
                    " possible.\n"
                ),
            },
            "dense": {
                "system_message": "You are a top university professor.",
                "instruction": (
                    "You are a top university professor. You have the below text and"
                    " you want to test the student's understanding of it. If you can"
                    " only ask {num_questions} question(s) but must cover all of the"
                    " content and the answer(s) to those questions must contain"
                    " **solely** the information presented in the given text, what"
                    " would you ask?\n"
                ),
            },
        }

    @staticmethod
    def sanitize_output_format(output: dict | list) -> list[dict]:
        """This static method takes in an `output` of type `dict` or `list` and returns
            a sanitized `list[dict]` output.

        Args:
            output (dict | list): The input `output` that needs to be sanitized.

        Returns:
            list[dict]: The sanitized output as a list of dictionaries.

        Raises:
            ValueError: If the `output` format is invalid.

        """
        if isinstance(output, dict):
            if list(output.keys()) == ["questions"]:
                if isinstance(output["questions"], list):
                    output = output["questions"]
            else:
                output = [output]
        elif isinstance(output, list):
            if list(output[0].keys()) == ["questions"]:
                output = output[0]["questions"]
        else:
            raise ValueError(f"Invalid output format: {type(output)}")

        return output  # type: ignore[return-value]

    def process(
        self,
        document: str,
        temperature: float = 1.0,
        question_type: str = "dense",
        num_questions: int = 5,
    ) -> tuple[list[dict[str, str]], list[dict]]:
        """
        Process the given document to generate a list of questions and answers.

        Args:
            document (str): The text document to process.
            temperature (float, optional): The temperature parameter for controlling
                the randomness of the output. Defaults to 1.0.
            question_type (str, optional): The type of questions to generate.
                Defaults to "dense".
            num_questions (int, optional): The number of questions to generate.
                Defaults to 5.

        Returns:
            tuple[list[dict[str, str]], list[dict]]: A tuple containing a list of
                questions and answers and a list of metadata.

        Raises:
            ValueError: If an invalid question type is provided.
        """
        if question_type not in self.question_types:
            raise ValueError(f"Invalid question type: {question_type}")

        system_message = self.question_types[question_type]["system_message"]
        instruction = self.question_types[question_type]["instruction"].format(
            num_questions=num_questions
        )
        messages = [
            {"role": "system", "content": system_message},
            {
                "role": "user",
                "content": instruction + self.output_format,
            },
            {"role": "user", "content": "\nHere is the given text:\n\n" + document},
        ]

        response = self.openai_client.chat.completions.create(
            model=self.openai_model,
            messages=messages,
            temperature=temperature,
            seed=self.seed,
            response_format={"type": "json_object"},
        )
        text = response.choices[0].message.content.strip()

        output = json.loads(text)
        output = self.sanitize_output_format(output)

        metadata = {}
        metadata["finish_reason"] = response.choices[0].finish_reason
        metadata["usage"] = response.usage.model_dump()

        return output, metadata  # type: ignore[return-value]

sanitize_output_format staticmethod 

1
sanitize_output_format(output)

This static method takes in an output of type dict or list and returns a sanitized list[dict] output.

Parameters:

Name Type Description Default
output dict | list

The input output that needs to be sanitized.

 required

Returns:

Type Description
list[dict]

list[dict]: The sanitized output as a list of dictionaries.

Raises:

Type Description
ValueError

If the output format is invalid.
Source code in docqa/core/data_generation.py 
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
@staticmethod
def sanitize_output_format(output: dict | list) -> list[dict]:
    """This static method takes in an `output` of type `dict` or `list` and returns
        a sanitized `list[dict]` output.

    Args:
        output (dict | list): The input `output` that needs to be sanitized.

    Returns:
        list[dict]: The sanitized output as a list of dictionaries.

    Raises:
        ValueError: If the `output` format is invalid.

    """
    if isinstance(output, dict):
        if list(output.keys()) == ["questions"]:
            if isinstance(output["questions"], list):
                output = output["questions"]
        else:
            output = [output]
    elif isinstance(output, list):
        if list(output[0].keys()) == ["questions"]:
            output = output[0]["questions"]
    else:
        raise ValueError(f"Invalid output format: {type(output)}")

    return output  # type: ignore[return-value]

process 

1
process(document, temperature=1.0, question_type='dense', num_questions=5)

Process the given document to generate a list of questions and answers.

Parameters:

Name Type Description Default
document str

The text document to process.

 required
temperature float

The temperature parameter for controlling the randomness of the output. Defaults to 1.0.

 1.0
question_type str

The type of questions to generate. Defaults to "dense".

 'dense'
num_questions int

The number of questions to generate. Defaults to 5.

 5

Returns:

Type Description
tuple[list[dict[str, str]], list[dict]]

tuple[list[dict[str, str]], list[dict]]: A tuple containing a list of questions and answers and a list of metadata.

Raises:

Type Description
ValueError

If an invalid question type is provided.
Source code in docqa/core/data_generation.py 
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def process(
    self,
    document: str,
    temperature: float = 1.0,
    question_type: str = "dense",
    num_questions: int = 5,
) -> tuple[list[dict[str, str]], list[dict]]:
    """
    Process the given document to generate a list of questions and answers.

    Args:
        document (str): The text document to process.
        temperature (float, optional): The temperature parameter for controlling
            the randomness of the output. Defaults to 1.0.
        question_type (str, optional): The type of questions to generate.
            Defaults to "dense".
        num_questions (int, optional): The number of questions to generate.
            Defaults to 5.

    Returns:
        tuple[list[dict[str, str]], list[dict]]: A tuple containing a list of
            questions and answers and a list of metadata.

    Raises:
        ValueError: If an invalid question type is provided.
    """
    if question_type not in self.question_types:
        raise ValueError(f"Invalid question type: {question_type}")

    system_message = self.question_types[question_type]["system_message"]
    instruction = self.question_types[question_type]["instruction"].format(
        num_questions=num_questions
    )
    messages = [
        {"role": "system", "content": system_message},
        {
            "role": "user",
            "content": instruction + self.output_format,
        },
        {"role": "user", "content": "\nHere is the given text:\n\n" + document},
    ]

    response = self.openai_client.chat.completions.create(
        model=self.openai_model,
        messages=messages,
        temperature=temperature,
        seed=self.seed,
        response_format={"type": "json_object"},
    )
    text = response.choices[0].message.content.strip()

    output = json.loads(text)
    output = self.sanitize_output_format(output)

    metadata = {}
    metadata["finish_reason"] = response.choices[0].finish_reason
    metadata["usage"] = response.usage.model_dump()

    return output, metadata  # type: ignore[return-value]

AnswerGenerator

Bases: BaseModel

Generate an answer to a question based on a reference.

Parameters:

Name Type Description Default
openai_key str

The OpenAI API key.

 required
openai_model str

The name of the OpenAI model to use.

 required
seed int

The seed for the random number generator. Defaults to 42.

 required
Source code in docqa/core/data_generation.py 
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
class AnswerGenerator(BaseModel):
    """Generate an answer to a question based on a reference.

    Args:
        openai_key (str): The OpenAI API key.
        openai_model (str): The name of the OpenAI model to use.
        seed (int, optional): The seed for the random number generator. Defaults to 42.
    """

    class Config:
        arbitrary_types_allowed = True

    openai_key: str
    openai_model: str
    seed: int = 42

    @computed_field  # type: ignore[misc]
    @property
    def openai_client(self) -> OpenAI:
        return OpenAI(api_key=self.openai_key)

    system_message: str = (
        "You are a trusted factual chatbot. You always answer questions based strictly"
        " on the provided reference."
    )
    instruction: str = (
        "Reference(s):\n\n{reference}\n\nStrictly according to the provided"
        " reference(s), give an answer as detailed as possible to the following"
        " question: {question}"
    )

    def process(
        self,
        question: str,
        reference: str,
        temperature: float = 1.0,
    ) -> tuple[str, dict]:
        """
        Process the given question and generate a response using the OpenAI model.

        Parameters:
            question (str): The question to be processed.
            reference (str): The reference string for the instruction.
            temperature (float, optional): The temperature parameter for generating the
                response. Higher values (e.g., 1.0) make the output more random, while
                lower values (e.g., 0.2) make it more focused and deterministic.
                Defaults to 1.0.

        Returns:
            Tuple[str, dict]: A tuple containing the generated answer and metadata.

        Output dict structure:
            - answer (str): The generated answer as a string.
            - metadata (dict): Additional metadata about the response.
                - finish_reason (str): The reason why the completion finished.
                - usage (dict): Usage statistics of the completion.
                    - completed_tokens (int): The number of tokens used for
                        completion.
                    - prompt_tokens (int): The number of tokens used for the prompt.
                    - total_tokens (int): The total number of tokens used.
        """
        messages = [
            {"role": "system", "content": self.system_message},
            {
                "role": "user",
                "content": self.instruction.format(
                    reference=reference, question=question
                ),
            },
        ]

        response = self.openai_client.chat.completions.create(
            model=self.openai_model,
            messages=messages,
            temperature=temperature,
            seed=self.seed,
        )
        answer = response.choices[0].message.content.strip()

        metadata = {}
        metadata["finish_reason"] = response.choices[0].finish_reason
        metadata["usage"] = {
            "completed_tokens": response.usage.completion_tokens,
            "prompt_tokens": response.usage.prompt_tokens,
            "total_tokens": response.usage.total_tokens,
        }

        return answer, metadata

process 

1
process(question, reference, temperature=1.0)

Process the given question and generate a response using the OpenAI model.

Parameters:

Name Type Description Default
question str

The question to be processed.

 required
reference str

The reference string for the instruction.

 required
temperature float

The temperature parameter for generating the response. Higher values (e.g., 1.0) make the output more random, while lower values (e.g., 0.2) make it more focused and deterministic. Defaults to 1.0.

 1.0

Returns:

Type Description
tuple[str, dict]

Tuple[str, dict]: A tuple containing the generated answer and metadata.
Output dict structure
  • answer (str): The generated answer as a string.
  • metadata (dict): Additional metadata about the response.
    • finish_reason (str): The reason why the completion finished.
    • usage (dict): Usage statistics of the completion.
      • completed_tokens (int): The number of tokens used for completion.
      • prompt_tokens (int): The number of tokens used for the prompt.
      • total_tokens (int): The total number of tokens used.
Source code in docqa/core/data_generation.py 
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
def process(
    self,
    question: str,
    reference: str,
    temperature: float = 1.0,
) -> tuple[str, dict]:
    """
    Process the given question and generate a response using the OpenAI model.

    Parameters:
        question (str): The question to be processed.
        reference (str): The reference string for the instruction.
        temperature (float, optional): The temperature parameter for generating the
            response. Higher values (e.g., 1.0) make the output more random, while
            lower values (e.g., 0.2) make it more focused and deterministic.
            Defaults to 1.0.

    Returns:
        Tuple[str, dict]: A tuple containing the generated answer and metadata.

    Output dict structure:
        - answer (str): The generated answer as a string.
        - metadata (dict): Additional metadata about the response.
            - finish_reason (str): The reason why the completion finished.
            - usage (dict): Usage statistics of the completion.
                - completed_tokens (int): The number of tokens used for
                    completion.
                - prompt_tokens (int): The number of tokens used for the prompt.
                - total_tokens (int): The total number of tokens used.
    """
    messages = [
        {"role": "system", "content": self.system_message},
        {
            "role": "user",
            "content": self.instruction.format(
                reference=reference, question=question
            ),
        },
    ]

    response = self.openai_client.chat.completions.create(
        model=self.openai_model,
        messages=messages,
        temperature=temperature,
        seed=self.seed,
    )
    answer = response.choices[0].message.content.strip()

    metadata = {}
    metadata["finish_reason"] = response.choices[0].finish_reason
    metadata["usage"] = {
        "completed_tokens": response.usage.completion_tokens,
        "prompt_tokens": response.usage.prompt_tokens,
        "total_tokens": response.usage.total_tokens,
    }

    return answer, metadata

generate_top_sections_questions 

1
generate_top_sections_questions(doc_tree, output_file, openai_key='', openai_model='', seed=42, temperature=1.0)

Generate the top sections with questions based on the provided document tree.

Parameters:

Name Type Description Default
doc_tree dict

The document tree representing the sections of the document.

 required
output_file Path

The path to the output file where the top sections with uestions will be saved.

 required
openai_key str

The OpenAI API key. Defaults to an empty string.

 ''
openai_model str

The OpenAI model to use for question generation. Defaults to an empty string.

 ''
seed int

The seed value for random number generation. Defaults to 42.

 42
temperature float

The temperature parameter for question generation. Defaults to 1.0.

 1.0

Returns:

Name Type Description
dict dict

The top sections with questions.
Source code in docqa/core/data_generation.py 
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
def generate_top_sections_questions(
    doc_tree: dict,
    output_file: Path,
    openai_key: str = "",
    openai_model: str = "",
    seed: int = 42,
    temperature: float = 1.0,
) -> dict:
    """
    Generate the top sections with questions based on the provided document tree.

    Args:
        doc_tree (dict): The document tree representing the sections of the document.
        output_file (Path): The path to the output file where the top sections with
            uestions will be saved.
        openai_key (str, optional): The OpenAI API key. Defaults to an empty string.
        openai_model (str, optional): The OpenAI model to use for question generation.
            Defaults to an empty string.
        seed (int, optional): The seed value for random number generation.
            Defaults to 42.
        temperature (float, optional): The temperature parameter for question
            generation.
            Defaults to 1.0.

    Returns:
        dict: The top sections with questions.
    """
    output_file = Path(output_file)
    if output_file.exists():
        with open(output_file, "r", encoding="utf-8") as f:
            top_sections_with_questions = json.load(f)
        return top_sections_with_questions

    qa_gen = QAPairGenerator(
        openai_key=openai_key,
        openai_model=openai_model,
        seed=seed,
    )

    top_sections_with_questions = {}
    if doc_tree["text"]:
        top_sections_with_questions[""] = {
            "text": doc_tree["text"],
            "chunks_count": len(chunk_content(doc_tree["text"])),
        }

    for section in doc_tree.get("child_sections", []):
        full_text = get_section_full_text(section)
        top_sections_with_questions[section["heading"]] = {
            "text": full_text,
            "chunks_count": len(chunk_content(full_text)),
        }

    for heading, section in top_sections_with_questions.items():
        print(f"Generating questions for {heading}")
        dense_questions, _ = qa_gen.process(
            section["text"],
            question_type="dense",
            num_questions=section["chunks_count"],
            temperature=temperature,
        )
        sparse_questions, _ = qa_gen.process(
            section["text"],
            question_type="sparse",
            temperature=temperature,
        )
        top_sections_with_questions[heading]["dense_questions"] = dense_questions
        top_sections_with_questions[heading]["sparse_questions"] = sparse_questions

    with open(output_file, "w", encoding="utf-8") as f:
        json.dump(top_sections_with_questions, f, indent=4, ensure_ascii=False)

    return top_sections_with_questions

generate_long_answers_for_sections_questions 

1
generate_long_answers_for_sections_questions(sections_with_questions, output_file, openai_key='', openai_model='', seed=42, temperature=1.0)

Generate long answers for sections' questions.

Parameters:

Name Type Description Default
sections_with_questions dict

A dictionary containing sections with their corresponding questions.

 required
output_file Path

The path to the output file where the generated long answers will be stored.

 required
openai_key str

The API key for OpenAI. Defaults to an empty string.

 ''
openai_model str

The name of the OpenAI model to use. Defaults to an empty string.

 ''
seed int

The seed value for random number generation. Defaults to 42.

 42
temperature float

The temperature parameter for generating answers. Defaults to 1.0.

 1.0

Returns:

Name Type Description
dict dict

A dictionary containing sections with their corresponding questions and generated long answers.
Source code in docqa/core/data_generation.py 
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
def generate_long_answers_for_sections_questions(
    sections_with_questions: dict,
    output_file: Path,
    openai_key: str = "",
    openai_model: str = "",
    seed: int = 42,
    temperature: float = 1.0,
) -> dict:
    """
    Generate long answers for sections' questions.

    Args:
        sections_with_questions (dict): A dictionary containing sections with their
            corresponding questions.
        output_file (Path): The path to the output file where the generated long answers
            will be stored.
        openai_key (str, optional): The API key for OpenAI. Defaults to an empty string.
        openai_model (str, optional): The name of the OpenAI model to use. Defaults to
            an empty string.
        seed (int, optional): The seed value for random number generation.
            Defaults to 42.
        temperature (float, optional): The temperature parameter for generating answers.
            Defaults to 1.0.

    Returns:
        dict: A dictionary containing sections with their corresponding questions and
            generated long answers.
    """
    output_file = Path(output_file)
    if output_file.exists():
        with open(output_file, "r", encoding="utf-8") as f:
            sections_with_questions_and_long_answers = json.load(f)
        return sections_with_questions_and_long_answers

    answer_gen = AnswerGenerator(
        openai_key=openai_key,
        openai_model=openai_model,
        seed=seed,
    )
    for heading, section in sections_with_questions.items():
        print(f"Generating long answers for dense questions of {heading}")
        reference = f"===\n[source: {heading}]\n{section['text']}\n===\n"

        for question in section["dense_questions"]:
            answer, _ = answer_gen.process(
                question=question["question"],
                reference=reference,
                temperature=temperature,
            )
            question["long_answer"] = answer

        print(f"Generating long answers for sparse questions of {heading}")
        for question in section["sparse_questions"]:
            answer, _ = answer_gen.process(
                question=question["question"],
                reference=reference,
                temperature=temperature,
            )
            question["long_answer"] = answer

    with open(output_file, "w", encoding="utf-8") as f:
        json.dump(sections_with_questions, f, indent=4, ensure_ascii=False)

    return sections_with_questions

make_simple_sample_for_openai 

1
make_simple_sample_for_openai(question, answer)

Generates a simple sample for OpenAI chat conversation.

Parameters:

Name Type Description Default
question str

The user's question.

 required
answer str

The assistant's answer.

 required

Returns:

Name Type Description
dict dict

A dictionary containing the chat conversation sample.
Example

make_simple_sample_for_openai("What is the capital of France?", "Paris")

Source code in docqa/core/data_generation.py 
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
def make_simple_sample_for_openai(question: str, answer: str) -> dict:
    """
    Generates a simple sample for OpenAI chat conversation.

    Args:
        question (str): The user's question.
        answer (str): The assistant's answer.

    Returns:
        dict: A dictionary containing the chat conversation sample.

    Example:
        make_simple_sample_for_openai("What is the capital of France?", "Paris")
    """
    return {
        "messages": [
            {
                "role": "system",
                "content": (
                    "You are a trusted factual chatbot that only answers questions"
                    " about generative agents."
                ),
            },
            {"role": "user", "content": question},
            {"role": "assistant", "content": answer},
        ]
    }

make_instruction_sample_for_openai 

1
make_instruction_sample_for_openai(question, answer, references)

Generates an instruction sample for OpenAI chat conversation.

Parameters:

Name Type Description Default
question str

The question to be used in the instruction.

 required
answer str

The answer to be used in the instruction.

 required
references list[str]

A list of reference texts to be included in the instruction.

 required

Returns:

Name Type Description
dict dict

A dictionary containing the chat conversation sample.
Source code in docqa/core/data_generation.py 
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
def make_instruction_sample_for_openai(
    question: str, answer: str, references: list[str]
) -> dict:
    """
    Generates an instruction sample for OpenAI chat conversation.

    Args:
        question (str): The question to be used in the instruction.
        answer (str): The answer to be used in the instruction.
        references (list[str]): A list of reference texts to be included in the
            instruction.

    Returns:
        dict: A dictionary containing the chat conversation sample.
    """
    reference_text = "\n\n".join(["===\n" + ref + "\n===" for ref in references])
    system_message = AnswerGenerator.model_fields["system_message"].default
    instruction = AnswerGenerator.model_fields["instruction"].default
    instruction = instruction.format(reference=reference_text, question=question)
    return {
        "messages": [
            {
                "role": "system",
                "content": system_message,
            },
            {
                "role": "user",
                "content": instruction,
            },
            {"role": "assistant", "content": answer},
        ]
    }