// Guide developers through creating a custom retrieval pipeline plugin for AutoRAG-Research. Walks through scaffolding, implementing BaseRetrievalPipeline methods, writing YAML configs, testing, and installing. Use when building a new search/retrieval strategy (e.g., Elasticsearch, ColBERT, custom vector search).
Guide developers through creating a custom generation pipeline plugin for AutoRAG-Research. Walks through scaffolding, implementing BaseGenerationPipeline methods, composing with retrieval pipelines, writing YAML configs, testing, and installing. Use when building a new RAG generation strategy (e.g., chain-of-thought RAG, multi-hop RAG).
Guide developers through creating a custom data ingestor plugin for AutoRAG-Research. Ingestors load external datasets (HuggingFace, local files, APIs) into the database. Uses @register_ingestor decorator for automatic CLI parameter extraction. Use when ingesting a new dataset format into AutoRAG-Research.
Guide developers through creating a custom evaluation metric plugin for AutoRAG-Research. Covers both retrieval metrics (recall, precision, etc.) and generation metrics (BLEU, ROUGE, etc.). Walks through scaffolding, implementing metric functions with @metric decorators, writing configs, testing, and installing. Use when building a new evaluation metric.
Query AutoRAG-Research pipeline results using natural language. Converts questions to SQL, executes safely (SELECT-only), returns formatted results. Auto-detects DB connection from configs/db.yaml or env vars. Use for pipeline comparison, metrics analysis, token usage.
Process [APPROVE] and [IGNORE] replies on /refactor review threads. Applies approved fixes to the codebase, resolves all responded threads on GitHub, commits and pushes changes. Sequential single-agent workflow. All output is in English.
Orchestrate a 3-agent PR code review debate using Claude Code Teams. Spawns Devil's Advocate, Neutral Judge, and Approval Advocate reviewers who analyze the current PR diff in parallel. Synthesizes findings, auto-fixes unanimous issues, and posts inline PR comments for disagreements. All output is in English.
| name | create-retrieval-plugin |
| description | Guide developers through creating a custom retrieval pipeline plugin for AutoRAG-Research. Walks through scaffolding, implementing BaseRetrievalPipeline methods, writing YAML configs, testing, and installing. Use when building a new search/retrieval strategy (e.g., Elasticsearch, ColBERT, custom vector search). |
| allowed-tools | ["Bash","Read","Write","Edit"] |
autorag-research plugin create my_search --type=retrieval
Read the generated pipeline.py, pyproject.toml, YAML config, and test file to understand the structure.
For the shared pipeline implementation and testing rules, read:
ai_instructions/pipeline_implementer.mdai_instructions/pipeline_test_writer.mdai_instructions/pipeline_architecture_mapper.mdImplement the two abstract methods in the pipeline class:
_retrieve_by_id(query_id, top_k) — retrieve using query ID (query exists in DB with stored embedding)_retrieve_by_text(query_text, top_k) — retrieve using raw text (may need on-the-fly embedding)Both must return list[dict[str, Any]] with doc_id (chunk ID) and score keys.
DO NOT add your own
asyncio.gather,asyncio.Semaphore, or any concurrency control. The base pipeline'srun()already handles parallel execution of all queries viarun_with_concurrency_limit()(semaphore + gather), controlled by themax_concurrencyconfig parameter. Your method is called once per single query — just implement the retrieval logic for that one query.
Custom parameters: Add fields to your config class and pass them via get_pipeline_kwargs() → accept them in the pipeline constructor. See bm25.py for a real example.
cd my_search_plugin
pip install -e . # or: uv pip install -e .
cd .. && autorag-research plugin sync
Verify: ls configs/pipelines/retrieval/my_search.yaml
| Purpose | Path |
|---|---|
| Base config class | autorag_research/config.py → BaseRetrievalPipelineConfig |
| Base pipeline class | autorag_research/pipelines/retrieval/base.py → BaseRetrievalPipeline |
| Service layer | autorag_research/orm/service/retrieval_pipeline.py → RetrievalPipelineService |
| Plugin entry point discovery | autorag_research/plugin_registry.py |
Study these existing implementations for patterns:
autorag_research/pipelines/retrieval/bm25.py — BM25 retrieval (simple)autorag_research/pipelines/retrieval/vector_search.py — Vector similarity searchautorag_research/pipelines/retrieval/hybrid.py — Hybrid (BM25 + vector)autorag_research/pipelines/retrieval/hyde.py — HyDE (Hypothetical Document Embeddings)configs/pipelines/retrieval/bm25.yaml, configs/pipelines/retrieval/vector_search.yaml