"""Dataset generation from documents.""" from __future__ import annotations import asyncio import re from typing import List from llama_index import Document, ServiceContext, SummaryIndex from llama_index.async_utils import DEFAULT_NUM_WORKERS, run_jobs from llama_index.core.response.schema import RESPONSE_TYPE from llama_index.ingestion import run_transformations from llama_index.llama_dataset import ( CreatedBy, CreatedByType, LabelledRagDataExample, LabelledRagDataset, ) from llama_index.postprocessor.node import KeywordNodePostprocessor from llama_index.prompts.base import BasePromptTemplate, PromptTemplate from llama_index.prompts.default_prompts import DEFAULT_TEXT_QA_PROMPT from llama_index.prompts.mixin import PromptDictType, PromptMixin, PromptMixinType from llama_index.schema import BaseNode, MetadataMode, NodeWithScore DEFAULT_QUESTION_GENERATION_PROMPT = """\ Context information is below. --------------------- {context_str} --------------------- Given the context information and not prior knowledge. generate only questions based on the below query. {query_str} """ class RagDatasetGenerator(PromptMixin): """Generate dataset (question/ question-answer pairs) \ based on the given documents. NOTE: this is a beta feature, subject to change! Args: nodes (List[Node]): List of nodes. (Optional) service_context (ServiceContext): Service Context. num_questions_per_chunk: number of question to be \ generated per chunk. Each document is chunked of size 512 words. text_question_template: Question generation template. question_gen_query: Question generation query. """ def __init__( self, nodes: List[BaseNode], service_context: ServiceContext | None = None, num_questions_per_chunk: int = 3, text_question_template: BasePromptTemplate | None = None, text_qa_template: BasePromptTemplate | None = None, question_gen_query: str | None = None, metadata_mode: MetadataMode = MetadataMode.NONE, show_progress: bool = False, workers: int = DEFAULT_NUM_WORKERS, ) -> None: """Init params.""" if service_context is None: service_context = service_context or ServiceContext.from_defaults( chunk_size_limit=3000 ) self.service_context = service_context self.text_question_template = text_question_template or PromptTemplate( DEFAULT_QUESTION_GENERATION_PROMPT ) self.text_qa_template = text_qa_template or DEFAULT_TEXT_QA_PROMPT self.question_gen_query = ( question_gen_query or f"You are a Teacher/Professor. Your task is to setup {num_questions_per_chunk} questions for an upcoming quiz/examination. The questions should be diverse in nature across the document. Restrict the questions to the context information provided." ) self.nodes = nodes self._metadata_mode = metadata_mode self._show_progress = show_progress self._workers = workers @classmethod def from_documents( cls, documents: List[Document], service_context: ServiceContext | None = None, num_questions_per_chunk: int = 3, text_question_template: BasePromptTemplate | None = None, text_qa_template: BasePromptTemplate | None = None, question_gen_query: str | None = None, required_keywords: List[str] | None = None, exclude_keywords: List[str] | None = None, show_progress: bool = False, workers: int = DEFAULT_NUM_WORKERS, ) -> RagDatasetGenerator: """Generate dataset from documents.""" if service_context is None: service_context = service_context or ServiceContext.from_defaults( chunk_size_limit=3000 ) nodes = run_transformations( documents, service_context.transformations, show_progress=show_progress ) # use node postprocessor to filter nodes required_keywords = required_keywords or [] exclude_keywords = exclude_keywords or [] node_postprocessor = KeywordNodePostprocessor( service_context=service_context, required_keywords=required_keywords, exclude_keywords=exclude_keywords, ) node_with_scores = [NodeWithScore(node=node) for node in nodes] node_with_scores = node_postprocessor.postprocess_nodes(node_with_scores) nodes = [node_with_score.node for node_with_score in node_with_scores] return cls( nodes=nodes, service_context=service_context, num_questions_per_chunk=num_questions_per_chunk, text_question_template=text_question_template, text_qa_template=text_qa_template, question_gen_query=question_gen_query, show_progress=show_progress, workers=workers, ) async def _agenerate_dataset( self, nodes: List[BaseNode], labelled: bool = False, ) -> LabelledRagDataset: """Node question generator.""" query_tasks = [] examples: List[LabelledRagDataExample] = [] summary_indices: List[SummaryIndex] = [] for node in nodes: index = SummaryIndex.from_documents( [ Document( text=node.get_content(metadata_mode=self._metadata_mode), metadata=node.metadata, excluded_llm_metadata_keys=node.excluded_llm_metadata_keys, excluded_embed_metadata_keys=node.excluded_embed_metadata_keys, relationships=node.relationships, ) ], service_context=self.service_context, ) query_engine = index.as_query_engine( service_context=self.service_context, text_qa_template=self.text_question_template, use_async=True, ) task = query_engine.aquery( self.question_gen_query, ) query_tasks.append(task) summary_indices.append(index) responses = await run_jobs(query_tasks, self._show_progress, self._workers) for idx, response in enumerate(responses): result = str(response).strip().split("\n") cleaned_questions = [ re.sub(r"^\d+[\).\s]", "", question).strip() for question in result ] cleaned_questions = [ question for question in cleaned_questions if len(question) > 0 ] index = summary_indices[idx] reference_context = nodes[idx].text model_name = self.service_context.llm.metadata.model_name created_by = CreatedBy(type=CreatedByType.AI, model_name=model_name) if labelled: index = summary_indices[idx] qr_tasks = [] for query in cleaned_questions: # build summary index off of node (i.e. context) qa_query_engine = index.as_query_engine( service_context=self.service_context, text_qa_template=self.text_qa_template, ) qr_task = qa_query_engine.aquery(query) qr_tasks.append(qr_task) answer_responses: List[RESPONSE_TYPE] = await run_jobs( qr_tasks, self._show_progress, self._workers ) for question, answer_response in zip( cleaned_questions, answer_responses ): example = LabelledRagDataExample( query=question, reference_answer=str(answer_response), reference_contexts=[reference_context], reference_answer_by=created_by, query_by=created_by, ) examples.append(example) else: for query in cleaned_questions: example = LabelledRagDataExample( query=query, reference_answer="", reference_contexts=[reference_context], reference_answer_by=None, query_by=created_by, ) examples.append(example) # split train/test return LabelledRagDataset(examples=examples) async def agenerate_questions_from_nodes(self) -> LabelledRagDataset: """Generates questions but not the reference answers.""" return await self._agenerate_dataset(self.nodes, labelled=False) async def agenerate_dataset_from_nodes(self) -> LabelledRagDataset: """Generates questions for each document.""" return await self._agenerate_dataset(self.nodes, labelled=True) def generate_questions_from_nodes(self) -> LabelledRagDataset: """Generates questions but not the reference answers.""" return asyncio.run(self.agenerate_questions_from_nodes()) def generate_dataset_from_nodes(self) -> LabelledRagDataset: """Generates questions for each document.""" return asyncio.run(self.agenerate_dataset_from_nodes()) def _get_prompts(self) -> PromptDictType: """Get prompts.""" return { "text_question_template": self.text_question_template, "text_qa_template": self.text_qa_template, } def _get_prompt_modules(self) -> PromptMixinType: """Get prompt modules.""" return {} def _update_prompts(self, prompts: PromptDictType) -> None: """Update prompts.""" if "text_question_template" in prompts: self.text_question_template = prompts["text_question_template"] if "text_qa_template" in prompts: self.text_qa_template = prompts["text_qa_template"]