Retrieval Synthetic Data Generation¶

Generate synthetic data for retrieval with two pipelines built on Data Designer. Text QA generates question-answer (QA) pairs from text documents. Multimodal generates queries with supporting text passages, images, or both.

Choose a Pipeline¶

Use the output you need to choose the pipeline:

Aspect	Text QA	Multimodal
Input	Directory of text files, with optional document bundling.	Prepared JSON Lines (JSONL) units, or a directory of independent text and image files.
Generation	Extract document artifacts, generate question-answer pairs, deduplicate questions, and score quality.	Plan contexts, generate queries, judge relevance, and identify supporting source units.
Result	Generated JSONL; convert it to retrieval training and evaluation files.	Portable retrieval bundle with corpus, queries, graded positives, and export views.
Command	`data-designer-retrieval-sdg generate`, then `convert`.	`python -m data_designer_retrieval_sdg.multimodal --config sdg.yaml`.
Python API	Imports from `data_designer_retrieval_sdg`.	Imports from `data_designer_retrieval_sdg.multimodal`.

The text QA pipeline retains its existing functionality, configuration, and package-root API. Its implementation now lives in text_qa/. Existing text QA jobs and recipe integrations can continue using the same commands and imports.

The multimodal pipeline is new functionality. It can also accept a text-only corpus, but produces query-to-source supervision rather than question-answer pairs. Its source format, configuration, output, and resume state are separate from text QA. It does not automatically replace a text QA job.

Start With Your Task¶

Choose a guide for the result you want:

Goal	Guide
Generate question-answer pairs from text documents	Generate Text QA Data
Try multimodal generation with a prepared NVIDIA hardware page	Create Your First Retrieval Bundle
Use your own text, image, or mixed corpus	Generate Retrieval Data From Text and Images
Look up multimodal inputs, model settings, and planning options	Multimodal Configuration Reference
Understand evidence, quality checks, views, and evaluation	Understand Multimodal Retrieval Data

Neither pipeline performs hard-negative mining or model training. Both pipelines filter generated data automatically. Optional sample review can help you refine generation for your corpus. Model-generated scores and labels can contain errors; use independent evaluation data to measure retrieval quality.

Data Designer Integration¶

Text QA builds one four-column Data Designer configuration and uses Data Designer's generation and resume support. Multimodal coordinates planning, model-call batches, evidence checks, and bundle export in Python. Each inference batch uses a registered Data Designer structured-response column.

The package registers three plugins automatically when installed:

Plugin	Role
`document-chunker`	Read and chunk text documents for text QA.
`embedding-dedup`	Deduplicate generated questions using text embeddings.
`retrieval-structured`	Decode and validate multimodal structured responses using Data Designer's generation loop.

Implementation and tests are organized under text_qa/ and multimodal/. The multimodal subpackage also owns its source loading, source models, query grouping, and split assignment.