Create Ingestor Plugin

Workflow

1. Scaffold

autorag-research plugin create my_dataset --type=ingestor

Read the generated ingestor.py, pyproject.toml, and test file to understand the structure.

The generated pyproject.toml registers the autorag_research.ingestors entry point. The @register_ingestor decorator handles automatic CLI parameter extraction from __init__ type hints.

2. Implement the ingestor

For the code-level implementation rules that are shared with the agent workflows, read:

• ai_instructions/implementation_specialist.md
• ai_instructions/schema_architect.md
• ai_instructions/test_writer.md

Required methods:

• __init__(embedding_model, ...) — accept embedding model + dataset-specific params
• detect_primary_key_type() → "bigint" or "string"
• ingest(subset, query_limit, min_corpus_cnt) — load data and save via self.service

__init__ type hints drive CLI generation automatically:

Parameters named embedding_model or late_interaction_embedding_model are auto-skipped (injected by CLI).

self.service is injected after construction via set_service(). Read existing ingestors for exact service method signatures.

3. Database Schema (critical)

Ingestors must populate the correct entity hierarchy:

Document → Page → Chunk (text)
                → ImageChunk (images)

• Document — top-level container (e.g., a Wikipedia article, a PDF)
• Page — subdivision within a document (linked via document_id)
• Chunk — text passage with embedding vector (linked to Page via PageChunkRelation)
• ImageChunk — image binary with embedding vector (linked to Page via PageChunkRelation)
• Query — search query with generation_gt: list[str] | None (ground truth answers)

RetrievalRelation — links queries to relevant chunks using AND/OR group structure:

RetrievalRelation(query_id, chunk_id, group_index, group_order, score)

group_index = AND group number
group_order = OR position within the group

Example: query needs (chunk_A OR chunk_B) AND chunk_C
  → (query, chunk_A, group_index=0, group_order=0)
  → (query, chunk_B, group_index=0, group_order=1)
  → (query, chunk_C, group_index=1, group_order=0)

This AND/OR structure is critical for multi-hop queries. See ai_instructions/db_schema.md for the full DBML schema.

4. Install and verify

cd my_dataset_plugin
pip install -e .   # or: uv pip install -e .

No plugin sync needed — ingestors are discovered automatically via entry points.

autorag-research ingest my_dataset --dataset-name subset_a

Testing

Use ingestor_test_utils for integration tests against a real PostgreSQL database:

• IngestorTestConfig — declare expected counts (queries, chunks, image_chunks), relation checks, primary key type
• create_test_database(config) — context manager that creates/drops an isolated test DB
• IngestorTestVerifier — runs all configured checks: count verification, format validation, retrieval relation checks, generation_gt checks, content hash verification

See tests/autorag_research/data/ingestor_test_utils.py for full API and usage examples in the module docstring.

Key Files

Examples

Study these existing implementations for patterns:

• autorag_research/data/beir.py — BEIR benchmark (simple, good starting point)
• autorag_research/data/bright.py — BRIGHT dataset
• autorag_research/data/mrtydi.py — Mr. TyDi multilingual dataset
• autorag_research/data/ragbench.py — RAGBench dataset